Search

Abstract Feature extraction and selection from medical data are the basis of radiomics and image biomarker discovery for various architectures, including convolutional neural networks (CNNs). We herein describe the typical radiomics steps and the components of a CNN for both deep feature extraction and end-to-end approaches. We discuss the curse of dimensionality, along with dimensionality reduction techniques. Despite the outstanding performance of deep learning (DL) approaches, the use of handcrafted features instead of deep learned features needs to be considered for each specific study. Dataset size is a key factor: large-scale datasets with low sample diversity could lead to overfitting; limited sample sizes can provide unstable models. The dataset must be representative of all the “facets” of the clinical phenomenon/disease investigated. The access to high-performance computational resources from graphics processing units is another key factor, especially for the training phase of deep architectures. The advantages of multi-institutional federated/collaborative learning are described. When large language models are used, high stability is needed to avoid catastrophic forgetting in complex domain-specific tasks. We highlight that non-DL approaches provide model explainability superior to that provided by DL approaches. To implement explainability, the need for explainable AI arises, also through post hoc mechanisms. Relevance statement This work aims to provide the key concepts for processing the imaging features to extract reliable and robust image biomarkers. Key Points The key concepts for processing the imaging features to extract reliable and robust image biomarkers are provided. The main differences between radiomics and representation learning approaches are highlighted. The advantages and disadvantages of handcrafted versus learned features are given without losing sight of the clinical purpose of artificial intelligence models. Graphical Abstract

Purpose: To assess radiomics and deep learning (DL) methods in identifying symptomatic Carotid Artery Disease (CAD) from carotid CT angiography (CTA) images. We further compare the performance of these novel methods to the conventional calcium score. Methods: Carotid CT angiography (CTA) images from symptomatic patients (ischaemic stroke/transient ischaemic attack within the last 3 months) and asymptomatic patients were analysed. Carotid arteries were classified into culprit, non-culprit and asymptomatic. The calcium score was assessed using the Agatston method. 93 radiomic features were extracted from regions-of-interest drawn on 14 consecutive CTA slices. For DL, convolutional neural networks (CNNs) with and without transfer learning were trained directly on CTA slices. Predictive performance was assessed over 5-fold cross validated AUC scores. SHAP and GRAD-CAM algorithms were used for explainability. Results: 132 carotid arteries were analysed (41 culprit, 41 non-culprit, and 50 asymptomatic). For asymptomatic vs symptomatic arteries, radiomics attained a mean AUC of 0.96(± 0.02), followed by DL 0.86(± 0.06) and then calcium 0.79(± 0.08). For culprit vs non-culprit arteries, radiomics achieved a mean AUC of 0.75(± 0.09), followed by DL 0.67(± 0.10) and then calcium 0.60(± 0.02). For multi-class classification, the mean AUCs were 0.95(± 0.07), 0.79(± 0.05), and 0.71(± 0.07) for radiomics, DL and calcium, respectively. Explainability revealed consistent patterns in the most important radiomic features. Conclusions: Our study highlights the potential of novel image analysis techniques in extracting quantitative information beyond calcification in the identification of CAD. Though further work is required, the transition of these novel techniques into clinical practice may eventually facilitate better stroke risk stratification.

Abstract Multiple sclerosis (MS) is a progressive demyelinating disease impacting the central nervous system. Conventional Magnetic Resonance Imaging (MRI) techniques (e.g., T2w images) help diagnose MS, although they sometimes reveal non-specific lesions. Quantitative MRI techniques are capable of quantifying imaging biomarkers in vivo, offering the potential to identify specific signs related to pre-clinical inflammation. Among those techniques, Quantitative Susceptibility Mapping (QSM) is particularly useful for studying processes that influence the magnetic properties of brain tissue, such as alterations in myelin concentration. Because of its intrinsic quantitative nature, it is particularly well-suited to be analyzed through radiomics, including techniques that extract a high number of complex and multi-dimensional features from radiological images. The dataset presented in this work provides information about normal-appearing white matter (NAWM) in a cohort of MS patients and healthy controls. It includes QSM-based radiomic features from NAWM and its tracts, and MR sequences necessary to implement the pipeline: T1w, T2w, QSM, DWI. The workflow is outlined in this article, along with an application showing feature reliability assessment.

Abstract Purpose To determine if pelvic/ovarian and omental lesions of ovarian cancer can be reliably segmented on computed tomography (CT) using fully automated deep learning-based methods. Methods A deep learning model for the two most common disease sites of high-grade serous ovarian cancer lesions (pelvis/ovaries and omentum) was developed and compared against the well-established “no-new-Net” framework and unrevised trainee radiologist segmentations. A total of 451 CT scans collected from four different institutions were used for training (n = 276), evaluation (n = 104) and testing (n = 71) of the methods. The performance was evaluated using the Dice similarity coefficient (DSC) and compared using a Wilcoxon test. Results Our model outperformed no-new-Net for the pelvic/ovarian lesions in cross-validation, on the evaluation and test set by a significant margin (p values being 4 × 10–7, 3 × 10–4, 4 × 10–2, respectively), and for the omental lesions on the evaluation set (p = 1 × 10–3). Our model did not perform significantly differently in segmenting pelvic/ovarian lesions (p = 0.371) compared to a trainee radiologist. On an independent test set, the model achieved a DSC performance of 71 ± 20 (mean ± standard deviation) for pelvic/ovarian and 61 ± 24 for omental lesions. Conclusion Automated ovarian cancer segmentation on CT scans using deep neural networks is feasible and achieves performance close to a trainee-level radiologist for pelvic/ovarian lesions. Relevance statement Automated segmentation of ovarian cancer may be used by clinicians for CT-based volumetric assessments and researchers for building complex analysis pipelines. Key points • The first automated approach for pelvic/ovarian and omental ovarian cancer lesion segmentation on CT images has been presented. • Automated segmentation of ovarian cancer lesions can be comparable with manual segmentation of trainee radiologists. • Careful hyperparameter tuning can provide models significantly outperforming strong state-of-the-art baselines. Graphical Abstract

Abstract High grade serous ovarian carcinoma (HGSOC) is a highly heterogeneous disease that typically presents at an advanced, metastatic state. The multi-scale complexity of HGSOC is a major obstacle to predicting response to neoadjuvant chemotherapy (NACT) and understanding critical determinants of response. Here we present a framework to predict the response of HGSOC patients to NACT integrating baseline clinical, blood-based, and radiomic biomarkers extracted from all primary and metastatic lesions. We use an ensemble machine learning model trained to predict the change in total disease volume using data obtained at diagnosis (n = 72). The model is validated in an internal hold-out cohort (n = 20) and an independent external patient cohort (n = 42). In the external cohort the integrated radiomics model reduces the prediction error by 8% with respect to the clinical model, achieving an AUC of 0.78 for RECIST 1.1 classification compared to 0.47 for the clinical model. Our results emphasize the value of including radiomics data in integrative models of treatment response and provide methods for developing new biomarker-based clinical trials of NACT in HGSOC.

Abstract Multiple Sclerosis (MS) is an autoimmune demyelinating disease characterised by changes in iron and myelin content. These biomarkers are detectable by Quantitative Susceptibility Mapping (QSM), an advanced Magnetic Resonance Imaging technique detecting magnetic properties. When analysed with radiomic techniques that exploit its intrinsic quantitative nature, QSM may furnish biomarkers to facilitate early diagnosis of MS and timely assessment of progression. In this work, we explore the robustness of QSM radiomic features by varying the number of grey levels (GLs) and echo times (TEs), in a sample of healthy controls and patients with MS. We analysed the white matter in total and within six clinically relevant tracts, including the cortico-spinal tract and the optic radiation. After optimising the number of GLs (n = 64), at least 65% of features were robust for each Volume of Interest (VOI), with no difference (p > .05) between left and right hemispheres. Different outcomes in feature robustness among the VOIs depend on their characteristics, such as volume and variance of susceptibility values. This study validated the processing pipeline for robustness analysis and established the reliability of QSM-based radiomics features against GLs and TEs. Our results provide important insights for future radiomics studies using QSM in clinical applications.

Deep learning (DL) networks have shown attractive performance in medical image processing tasks such as brain tumor classification. However, they are often criticized as mysterious “black boxes”. The opaqueness of the model and the reasoning process make it difficult for health workers to decide whether to trust the prediction outcomes. In this study, we develop an interpretable multi-part attention network (IMPA-Net) for brain tumor classification to enhance the interpretability and trustworthiness of classification outcomes. The proposed model not only predicts the tumor grade but also provides a global explanation for the model interpretability and a local explanation as justification for the proffered prediction. Global explanation is represented as a group of feature patterns that the model learns to distinguish high-grade glioma (HGG) and low-grade glioma (LGG) classes. Local explanation interprets the reasoning process of an individual prediction by calculating the similarity between the prototypical parts of the image and a group of pre-learned task-related features. Experiments conducted on the BraTS2017 dataset demonstrate that IMPA-Net is a verifiable model for the classification task. A percentage of 86% of feature patterns were assessed by two radiologists to be valid for representing task-relevant medical features. The model shows a classification accuracy of 92.12%, of which 81.17% were evaluated as trustworthy based on local explanations. Our interpretable model is a trustworthy model that can be used for decision aids for glioma classification. Compared with black-box CNNs, it allows health workers and patients to understand the reasoning process and trust the prediction outcomes.

Abstract Photoacoustic imaging is an increasingly popular method of exploring the tumour microenvironment, which can provide insight into tumour oxygenation status and potentially treatment response assessment. Currently, the measurements most commonly performed on such images are the mean and median of the pixel values of the tumour volumes of interest. We investigated expanding the set of measurements that can be extracted from these images by adding radiomic features. In particular, we found that Skewness was sensitive to differences between basal and luminal patient derived xenograft cancer models with an $$\eta ^2$$ η 2 of 0.86, and that it was robust to variations in confounding factors such as reconstruction type and wavelength. We also built discriminant models with radiomic features that were correlated with the underlying tumour model and were independent from each other. We then ranked features by their importance in the model. Skewness was again found to be an important feature, as were 10th Percentile, Root Mean Squared, and several other texture-based features. In summary, this paper proposes a methodology to select radiomic features extracted from photoacoustic images that are robust to changes in acquisition and reconstruction parameters, and discusses features found to have discriminating power between the underlying tumour models in a pre-clinical dataset.

Introduction: The aim of our study was to evaluate the feasibility of texture analysis of epicardial fat (EF) and thoracic subcutaneous fat (TSF) in patients undergoing cardiac CT (CCT). Materials and methods: We compared a consecutive population of 30 patients with BMI ≤25 kg/m2 (Group A, 60.6 ± 13.7 years) with a control population of 30 patients with BMI >25 kg/m2 (Group B, 63.3 ± 11 years). A dedicated computer application for quantification of EF and a texture analysis application for the study of EF and TSF were employed. Results: The volume of EF was higher in group B (mean 116.1 cm3 vs. 86.3 cm3, p = 0.014), despite no differences were found neither in terms of mean density (−69.5 ± 5 HU vs. −68 ± 5 HU, p = 0.28), nor in terms of quartiles distribution (Q1, p = 0.83; Q2, p = 0.22, Q3, p = 0.83, Q4, p = 0.34). The discriminating parameters of the histogram class were mean (p = 0.02), 0,1st (p = 0.001), 10th (p = 0.002), and 50th percentiles (p = 0.02). DifVarnc was the discriminating parameter of the co-occurrence matrix class (p = 0.007).The TSF thickness was 15 ± 6 mm in group A and 19.5 ± 5 mm in group B (p = 0.003). The TSF had a mean density of −97 ± 19 HU in group A and −95.8 ± 19 HU in group B (p = 0.75). The discriminating parameters of texture analysis were 10th (p = 0.03), 50th (p = 0.01), 90th percentiles (p = 0.04), S(0,1)SumAverg (p = 0.02), S(1,-1)SumOfSqs (p = 0.02), S(3,0)Contrast (p = 0.03), S(3,0)SumAverg (p = 0.02), S(4,0)SumAverg (p = 0.04), Horzl_RLNonUni (p = 0.02), and Vertl_LngREmph (p = 0.0005). Conclusions: Texture analysis provides distinctive radiomic parameters of EF and TSF. EF and TSF had different radiomic features as the BMI varies.

Abstract Objectives We systematically reviewed the current literature evaluating the ability of fully-automated deep learning (DL) and semi-automated traditional machine learning (TML) MRI-based artificial intelligence (AI) methods to differentiate clinically significant prostate cancer (csPCa) from indolent PCa (iPCa) and benign conditions. Methods We performed a computerised bibliographic search of studies indexed in MEDLINE/PubMed, arXiv, medRxiv, and bioRxiv between 1 January 2016 and 31 July 2021. Two reviewers performed the title/abstract and full-text screening. The remaining papers were screened by four reviewers using the Checklist for Artificial Intelligence in Medical Imaging (CLAIM) for DL studies and Radiomics Quality Score (RQS) for TML studies. Papers that fulfilled the pre-defined screening requirements underwent full CLAIM/RQS evaluation alongside the risk of bias assessment using QUADAS-2, both conducted by the same four reviewers. Standard measures of discrimination were extracted for the developed predictive models. Results 17/28 papers (five DL and twelve TML) passed the quality screening and were subject to a full CLAIM/RQS/QUADAS-2 assessment, which revealed a substantial study heterogeneity that precluded us from performing quantitative analysis as part of this review. The mean RQS of TML papers was 11/36, and a total of five papers had a high risk of bias. AUCs of DL and TML papers with low risk of bias ranged between 0.80–0.89 and 0.75–0.88, respectively. Conclusion We observed comparable performance of the two classes of AI methods and identified a number of common methodological limitations and biases that future studies will need to address to ensure the generalisability of the developed models.

BackgroundHigh-Grade Serous Ovarian Carcinoma (HGSOC) is the most prevalent and lethal subtype of ovarian cancer, but has a paucity of clinically-actionable biomarkers due to high degrees of multi-level heterogeneity. Radiogenomics markers have the potential to improve prediction of patient outcome and treatment response, but require accurate multimodal spatial registration between radiological imaging and histopathological tissue samples. Previously published co-registration work has not taken into account the anatomical, biological and clinical diversity of ovarian tumours.MethodsIn this work, we developed a research pathway and an automated computational pipeline to produce lesion-specific three-dimensional (3D) printed moulds based on preoperative cross-sectional CT or MRI of pelvic lesions. Moulds were designed to allow tumour slicing in the anatomical axial plane to facilitate detailed spatial correlation of imaging and tissue-derived data. Code and design adaptations were made following each pilot case through an iterative refinement process.ResultsFive patients with confirmed or suspected HGSOC who underwent debulking surgery between April and December 2021 were included in this prospective study. Tumour moulds were designed and 3D-printed for seven pelvic lesions, covering a range of tumour volumes (7 to 133 cm3) and compositions (cystic and solid proportions). The pilot cases informed innovations to improve specimen and subsequent slice orientation, through the use of 3D-printed tumour replicas and incorporation of a slice orientation slit in the mould design, respectively. The overall research pathway was compatible with implementation within the clinically determined timeframe and treatment pathway for each case, involving multidisciplinary clinical professionals from Radiology, Surgery, Oncology and Histopathology Departments.ConclusionsWe developed and refined a computational pipeline that can model lesion-specific 3D-printed moulds from preoperative imaging for a variety of pelvic tumours. This framework can be used to guide comprehensive multi-sampling of tumour resection specimens.

Abstract Robust machine learning models based on radiomic features might allow for accurate diagnosis, prognosis, and medical decision-making. Unfortunately, the lack of standardized radiomic feature extraction has hampered their clinical use. Since the radiomic features tend to be affected by low voxel statistics in regions of interest, increasing the sample size would improve their robustness in clinical studies. Therefore, we propose a Generative Adversarial Network (GAN)-based lesion-focused framework for Computed Tomography (CT) image Super-Resolution (SR); for the lesion (i.e., cancer) patch-focused training, we incorporate Spatial Pyramid Pooling (SPP) into GAN-Constrained by the Identical, Residual, and Cycle Learning Ensemble (GAN-CIRCLE). At $$2\times $$ 2 × SR, the proposed model achieved better perceptual quality with less blurring than the other considered state-of-the-art SR methods, while producing comparable results at $$4\times $$ 4 × SR. We also evaluated the robustness of our model’s radiomic feature in terms of quantization on a different lung cancer CT dataset using Principal Component Analysis (PCA). Intriguingly, the most important radiomic features in our PCA-based analysis were the most robust features extracted on the GAN-super-resolved images. These achievements pave the way for the application of GAN-based image Super-Resolution techniques for studies of radiomics for robust biomarker discovery.

The epicardial adipose tissue (EAT) is the visceral fat located between the myocardium and the pericardium. We aimed to perform a longitudinal evaluation of the epicardial adipose tissue using an advanced computer-assisted approach in a population of patients undergoing Cardiac CT (CCT) during statin administration, in relation to their body mass index (BMI). We retrospectively enrolled 95 patients [mean age 62 ± 10 years; 68 males (72%) and 27 females (28%)] undergoing CCT for suspected coronary artery disease during statin administration. CCT was performed at two subsequent time points. At the second CCT, EAT showed a mean density increase (−75.59 ± 7.0 HU vs. −78.18 ± 5.3 HU, p < 0.001) and a volume decrease (130 ± 54.3 cm3 vs.142.79 ± 56.9 cm3, p < 0.001). Concerning coronary artery EAT thickness, a reduction was found at the origin of the right coronary artery (13.26 ± 5.2 mm vs. 14.94 ± 5.8, p = 0.001) and interventricular artery (8.22 ± 3.7 mm vs. 9.13 ± 3.9 mm, p = 0.001). The quartile (Q) attenuation percentage (%) distribution of EAT changed at the second CCT. The EAT % distribution changed by the BMI in Q1 (p = 0.015), Q3 (p = 0.001) and Q4 (p = 0.043) at the second CCT, but the normal-BMI and overweight/obese patients showed a similar response to statin therapy in terms of quartile distribution changes. In conclusion, statins may determine significant changes in EAT quantitative and qualitative characteristics detected by CCT; the BMI influences the EAT composition, but statins determine a similar response in quartile distribution’s variation, irrespective of the BMI.

Artificial intelligence (AI) methods applied to healthcare problems have shown enormous potential to alleviate the burden of health services worldwide and to improve the accuracy and reproducibility of predictions. In particular, developments in computer vision are creating a paradigm shift in the analysis of radiological images, where AI tools are already capable of automatically detecting and precisely delineating tumours. However, such tools are generally developed in technical departments that continue to be siloed from where the real benefit would be achieved with their usage. Significant effort still needs to be made to make these advancements available, first in academic clinical research and ultimately in the clinical setting. In this paper, we demonstrate a prototype pipeline based entirely on open-source software and free of cost to bridge this gap, simplifying the integration of tools and models developed within the AI community into the clinical research setting, ensuring an accessible platform with visualisation applications that allow end-users such as radiologists to view and interact with the outcome of these AI tools.

Abstract Nearly half of patients with prostate cancer (PCa) harbour low- or intermediate-risk disease considered suitable for active surveillance (AS). However, up to 44% of patients discontinue AS within the first five years, highlighting the unmet clinical need for robust baseline risk-stratification tools that enable timely and accurate prediction of tumour progression. In this proof-of-concept study, we sought to investigate the added value of MRI-derived radiomic features to standard-of-care clinical parameters for improving baseline prediction of PCa progression in AS patients. Tumour T2-weighted imaging (T2WI) and apparent diffusion coefficient radiomic features were extracted, with rigorous calibration and pre-processing methods applied to select the most robust features for predictive modelling. Following leave-one-out cross-validation, the addition of T2WI-derived radiomic features to clinical variables alone improved the area under the ROC curve for predicting progression from 0.61 (95% confidence interval [CI] 0.481–0.743) to 0.75 (95% CI 0.64–0.86). These exploratory findings demonstrate the potential benefit of MRI-derived radiomics to add incremental benefit to clinical data only models in the baseline prediction of PCa progression on AS, paving the way for future multicentre studies validating the proposed model and evaluating its impact on clinical outcomes.

Abstract Computed Tomography (CT) is widely used in oncology for morphological evaluation and diagnosis, commonly through visual assessments, often exploiting semi-automatic tools as well. Well-established automatic methods for quantitative imaging offer the opportunity to enrich the radiologist interpretation with a large number of radiomic features, which need to be highly reproducible to be used reliably in clinical practice. This study investigates feature reproducibility against noise, varying resolutions and segmentations (achieved by perturbing the regions of interest), in a CT dataset with heterogeneous voxel size of 98 renal cell carcinomas (RCCs) and 93 contralateral normal kidneys (CK). In particular, first order (FO) and second order texture features based on both 2D and 3D grey level co-occurrence matrices (GLCMs) were considered. Moreover, this study carries out a comparative analysis of three of the most commonly used interpolation methods, which need to be selected before any resampling procedure. Results showed that the Lanczos interpolation is the most effective at preserving original information in resampling, where the median slice resolution coupled with the native slice spacing allows the best reproducibility, with 94.6% and 87.7% of features, in RCC and CK, respectively. GLCMs show their maximum reproducibility when used at short distances.

Abstract Background Unsupervised learning can discover various unseen abnormalities, relying on large-scale unannotated medical images of healthy subjects. Towards this, unsupervised methods reconstruct a 2D/3D single medical image to detect outliers either in the learned feature space or from high reconstruction loss. However, without considering continuity between multiple adjacent slices, they cannot directly discriminate diseases composed of the accumulation of subtle anatomical anomalies, such as Alzheimer’s disease (AD). Moreover, no study has shown how unsupervised anomaly detection is associated with either disease stages, various (i.e., more than two types of) diseases, or multi-sequence magnetic resonance imaging (MRI) scans. Results We propose unsupervised medical anomaly detection generative adversarial network (MADGAN), a novel two-step method using GAN-based multiple adjacent brain MRI slice reconstruction to detect brain anomalies at different stages on multi-sequence structural MRI: (Reconstruction) Wasserstein loss with Gradient Penalty + 100 $$\ell _1$$ ℓ 1 loss—trained on 3 healthy brain axial MRI slices to reconstruct the next 3 ones—reconstructs unseen healthy/abnormal scans; (Diagnosis) Average $$\ell _2$$ ℓ 2 loss per scan discriminates them, comparing the ground truth/reconstructed slices. For training, we use two different datasets composed of 1133 healthy T1-weighted (T1) and 135 healthy contrast-enhanced T1 (T1c) brain MRI scans for detecting AD and brain metastases/various diseases, respectively. Our self-attention MADGAN can detect AD on T1 scans at a very early stage, mild cognitive impairment (MCI), with area under the curve (AUC) 0.727, and AD at a late stage with AUC 0.894, while detecting brain metastases on T1c scans with AUC 0.921. Conclusions Similar to physicians’ way of performing a diagnosis, using massive healthy training data, our first multiple MRI slice reconstruction approach, MADGAN, can reliably predict the next 3 slices from the previous 3 ones only for unseen healthy images. As the first unsupervised various disease diagnosis, MADGAN can reliably detect the accumulation of subtle anatomical anomalies and hyper-intense enhancing lesions, such as (especially late-stage) AD and brain metastases on multi-sequence MRI scans.

Abstract Radiomic image features are becoming a promising non-invasive method to obtain quantitative measurements for tumour classification and therapy response assessment in oncological research. However, despite its increasingly established application, there is a need for standardisation criteria and further validation of feature robustness with respect to imaging acquisition parameters. In this paper, the robustness of radiomic features extracted from computed tomography (CT) images is evaluated for liver tumour and muscle, comparing the values of the features in images reconstructed with two different slice thicknesses of 2.0 mm and 5.0 mm. Novel approaches are presented to address the intrinsic dependencies of texture radiomic features, choosing the optimal number of grey levels and correcting for the dependency on volume. With the optimal values and corrections, feature values are compared across thicknesses to identify reproducible features. Normalisation using muscle regions is also described as an alternative approach. With either method, a large fraction of features (75–90%) was found to be highly robust (

Abstract Recent advances in Quantum Machine Learning (QML) have provided benefits to several computational processes, drastically reducing the time complexity. Another approach of combining quantum information theory with machine learning—without involving quantum computers—is known as Quantum-inspired Machine Learning (QiML), which exploits the expressive power of the quantum language to increase the accuracy of the process (rather than reducing the time complexity). In this work, we propose a large-scale experiment based on the application of a binary classifier inspired by quantum information theory to the biomedical imaging context in clonogenic assay evaluation to identify the most discriminative feature, allowing us to enhance cell colony segmentation. This innovative approach offers a two-fold result: (1) among the extracted and analyzed image features, homogeneity is shown to be a relevant feature in detecting challenging cell colonies; and (2) the proposed quantum-inspired classifier is a novel and outstanding methodology, compared to conventional machine learning classifiers, for the evaluation of clonogenic assays.

Abstract Radiomics, quantitative feature extraction from radiological images, can improve disease diagnosis and prognostication. However, radiomic features are susceptible to image acquisition and segmentation variability. Ideally, only features robust to these variations would be incorporated into predictive models, for good generalisability. We extracted 93 radiomic features from carotid artery computed tomography angiograms of 41 patients with cerebrovascular events. We tested feature robustness to region-of-interest perturbations, image pre-processing settings and quantisation methods using both single- and multi-slice approaches. We assessed the ability of the most robust features to identify culprit and non-culprit arteries using several machine learning algorithms and report the average area under the curve (AUC) from five-fold cross validation. Multi-slice features were superior to single for producing robust radiomic features (67 vs. 61). The optimal image quantisation method used bin widths of 25 or 30. Incorporating our top 10 non-redundant robust radiomics features into ElasticNet achieved an AUC of 0.73 and accuracy of 69% (compared to carotid calcification alone [AUC: 0.44, accuracy: 46%]). Our results provide key information for introducing carotid CT radiomics into clinical practice. If validated prospectively, our robust carotid radiomic set could improve stroke prediction and target therapies to those at highest risk.

Abstract The recent developments of information technologies, and the consequent need for access to distributed services and resources, require robust and reliable authentication systems. Biometric systems can guarantee high levels of security and multimodal techniques, which combine two or more biometric traits, warranting constraints that are more stringent during the access phases. This work proposes a novel multimodal biometric system based on iris and retina combination in the spatial domain. The proposed solution follows the alignment and recognition approach commonly adopted in computational linguistics and bioinformatics; in particular, features are extracted separately for iris and retina, and the fusion is obtained relying upon the comparison score via the Levenshtein distance. We evaluated our approach by testing several combinations of publicly available biometric databases, namely one for retina images and three for iris images. To provide comprehensive results, detection error trade‐off‐based metrics, as well as statistical analyses for assessing the authentication performance, were considered. The best achieved False Acceptation Rate and False Rejection Rate indices were and 3.33%, respectively, for the multimodal retina‐iris biometric approach that overall outperformed the unimodal systems. These results draw the potential of the proposed approach as a multimodal authentication framework using multiple static biometric traits.

BackgroundPathological response to neoadjuvant treatment for patients with high-grade serous ovarian carcinoma (HGSOC) is assessed using the chemotherapy response score (CRS) for omental tumor deposits. The main limitation of CRS is that it requires surgical sampling after initial neoadjuvant chemotherapy (NACT) treatment. Earlier and non-invasive response predictors could improve patient stratification. We developed computed tomography (CT) radiomic measures to predict neoadjuvant response before NACT using CRS as a gold standard.MethodsOmental CT-based radiomics models, yielding a simplified fully interpretable radiomic signature, were developed using Elastic Net logistic regression and compared to predictions based on omental tumor volume alone. Models were developed on a single institution cohort of neoadjuvant-treated HGSOC (n = 61; 41% complete response to NCT) and tested on an external test cohort (n = 48; 21% complete response).ResultsThe performance of the comprehensive radiomics models and the fully interpretable radiomics model was significantly higher than volume-based predictions of response in both the discovery and external test sets when assessed using G-mean (geometric mean of sensitivity and specificity) and NPV, indicating high generalizability and reliability in identifying non-responders when using radiomics. The performance of a fully interpretable model was similar to that of comprehensive radiomics models.ConclusionsCT-based radiomics allows for predicting response to NACT in a timely manner and without the need for abdominal surgery. Adding pre-NACT radiomics to volumetry improved model performance for predictions of response to NACT in HGSOC and was robust to external testing. A radiomic signature based on five robust predictive features provides improved clinical interpretability and may thus facilitate clinical acceptance and application.

Abstract Background Ovarian cancer survival rates have not changed in the last 20 years. The majority of cases are High-grade serous ovarian carcinomas (HGSOCs), which are typically diagnosed at an advanced stage with multiple metastatic lesions. Taking biopsies of all sites of disease is infeasible, which challenges the implementation of stratification tools based on molecular profiling. Main body In this review, we describe how these challenges might be overcome by integrating quantitative features extracted from medical imaging with the analysis of paired genomic profiles, a combined approach called radiogenomics, to generate virtual biopsies. Radiomic studies have been used to model different imaging phenotypes, and some radiomic signatures have been associated with paired molecular profiles to monitor spatiotemporal changes in the heterogeneity of tumours. We describe different strategies to integrate radiogenomic information in a global and local manner, the latter by targeted sampling of tumour habitats, defined as regions with distinct radiomic phenotypes. Conclusion Linking radiomics and biological correlates in a targeted manner could potentially improve the clinical management of ovarian cancer. Radiogenomic signatures could be used to monitor tumours during the course of therapy, offering additional information for clinical decision making. In summary, radiogenomics may pave the way to virtual biopsies and treatment monitoring tools for integrative tumour analysis.

Mathematical models of biochemical networks can largely facilitate the comprehension of the mechanisms at the basis of cellular processes, as well as the formulation of hypotheses that can be tested by means of targeted laboratory experiments. However, two issues might hamper the achievement of fruitful outcomes. On the one hand, detailed mechanistic models can involve hundreds or thousands of molecular species and their intermediate complexes, as well as hundreds or thousands of chemical reactions, a situation generally occurring in rule-based modeling. On the other hand, the computational analysis of a model typically requires the execution of a large number of simulations for its calibration, or to test the effect of perturbations. As a consequence, the computational capabilities of modern Central Processing Units can be easily overtaken, possibly making the modeling of biochemical networks a worthless or ineffective effort. To the aim of overcoming the limitations of the current state-of-the-art simulation approaches, we present in this paper FiCoS, a novel "black-box" deterministic simulator that effectively realizes both a fine-grained and a coarse-grained parallelization on Graphics Processing Units. In particular, FiCoS exploits two different integration methods, namely, the Dormand-Prince and the Radau IIA, to efficiently solve both non-stiff and stiff systems of coupled Ordinary Differential Equations. We tested the performance of FiCoS against different deterministic simulators, by considering models of increasing size and by running analyses with increasing computational demands. FiCoS was able to dramatically speedup the computations up to 855×, showing to be a promising solution for the simulation and analysis of large-scale models of complex biological processes.

Abstract Background In order to fully characterize the genome of an individual, the reconstruction of the two distinct copies of each chromosome, called haplotypes, is essential. The computational problem of inferring the full haplotype of a cell starting from read sequencing data is known as haplotype assembly, and consists in assigning all heterozygous Single Nucleotide Polymorphisms (SNPs) to exactly one of the two chromosomes. Indeed, the knowledge of complete haplotypes is generally more informative than analyzing single SNPs and plays a fundamental role in many medical applications. Results To reconstruct the two haplotypes, we addressed the weighted Minimum Error Correction (wMEC) problem, which is a successful approach for haplotype assembly. This NP-hard problem consists in computing the two haplotypes that partition the sequencing reads into two disjoint sub-sets, with the least number of corrections to the SNP values. To this aim, we propose here GenHap, a novel computational method for haplotype assembly based on Genetic Algorithms, yielding optimal solutions by means of a global search process. In order to evaluate the effectiveness of our approach, we run GenHap on two synthetic (yet realistic) datasets, based on the Roche/454 and PacBio RS II sequencing technologies. We compared the performance of GenHap against HapCol, an efficient state-of-the-art algorithm for haplotype phasing. Our results show that GenHap always obtains high accuracy solutions (in terms of haplotype error rate), and is up to 4× faster than HapCol in the case of Roche/454 instances and up to 20× faster when compared on the PacBio RS II dataset. Finally, we assessed the performance of GenHap on two different real datasets. Conclusions Future-generation sequencing technologies, producing longer reads with higher coverage, can highly benefit from GenHap, thanks to its capability of efficiently solving large instances of the haplotype assembly problem. Moreover, the optimization approach proposed in GenHap can be extended to the study of allele-specific genomic features, such as expression, methylation and chromatin conformation, by exploiting multi-objective optimization techniques. The source code and the full documentation are available at the following GitHub repository: https://github.com/andrea-tango/GenHap.

Convolutional Neural Networks (CNNs) achieve excellent computer-assisted diagnosis with sufficient annotated training data. However, most medical imaging datasets are small and fragmented. In this context, Generative Adversarial Networks (GANs) can synthesize realistic/diverse additional training images to fill the data lack in the real image distribution; researchers have improved classification by augmenting data with noise-to-image (e.g., random noise samples to diverse pathological images) or image-to-image GANs (e.g., a benign image to a malignant one). Yet, no research has reported results combining noise-to-image and image-to-image GANs for further performance boost. Therefore, to maximize the DA effect with the GAN combinations, we propose a two-step GAN-based DA that generates and refines brain Magnetic Resonance (MR) images with/without tumors separately: (i) Progressive Growing of GANs (PGGANs), multi-stage noise-to-image GAN for high-resolution MR image generation, first generates realistic/diverse 256×256 images; (ii) Multimodal UNsupervised Image-to-image Translation (MUNIT) that combines GANs/Variational AutoEncoders or SimGAN that uses a DA-focused GAN loss, further refines the texture/shape of the PGGAN-generated images similarly to the real ones. We thoroughly investigate CNN-based tumor classification results, also considering the influence of pre-training on ImageNet and discarding weird-looking GAN-generated images. The results show that, when combined with classic DA, our two-step GAN-based DA can significantly outperform the classic DA alone, in tumor detection (i.e., boosting sensitivity 93.67% to 97.48%) and also in other medical imaging tasks.

The Special Issue “Artificial Intelligence Applied to Medical Imaging and Computational Biology” of the Applied Sciences Journal has been curated from February 2021 to May 2022, which covered the state-of-the-art and novel algorithms and applications of Artificial Intelligence methods for biomedical data analysis, ranging from classic Machine Learning to Deep Learning [...]

Given growing clinical needs, in recent years Artificial Intelligence (AI) techniques have increasingly been used to define the best approaches for survival assessment and prediction in patients with brain tumors. Advances in computational resources, and the collection of (mainly) public databases, have promoted this rapid development. This narrative review of the current state-of-the-art aimed to survey current applications of AI in predicting survival in patients with brain tumors, with a focus on Magnetic Resonance Imaging (MRI). An extensive search was performed on PubMed and Google Scholar using a Boolean research query based on MeSH terms and restricting the search to the period between 2012 and 2022. Fifty studies were selected, mainly based on Machine Learning (ML), Deep Learning (DL), radiomics-based methods, and methods that exploit traditional imaging techniques for survival assessment. In addition, we focused on two distinct tasks related to survival assessment: the first on the classification of subjects into survival classes (short and long-term or eventually short, mid and long-term) to stratify patients in distinct groups. The second focused on quantification, in days or months, of the individual survival interval. Our survey showed excellent state-of-the-art methods for the first, with accuracy up to ∼98%. The latter task appears to be the most challenging, but state-of-the-art techniques showed promising results, albeit with limitations, with C-Index up to ∼0.91. In conclusion, according to the specific task, the available computational methods perform differently, and the choice of the best one to use is non-univocal and dependent on many aspects. Unequivocally, the use of features derived from quantitative imaging has been shown to be advantageous for AI applications, including survival prediction. This evidence from the literature motivates further research in the field of AI-powered methods for survival prediction in patients with brain tumors, in particular, using the wealth of information provided by quantitative MRI techniques.

Convolutional neural networks (CNNs) constitute a widely used deep learning approach that has frequently been applied to the problem of brain tumor diagnosis. Such techniques still face some critical challenges in moving towards clinic application. The main objective of this work is to present a comprehensive review of studies using CNN architectures to classify brain tumors using MR images with the aim of identifying useful strategies for and possible impediments in the development of this technology. Relevant articles were identified using a predefined, systematic procedure. For each article, data were extracted regarding training data, target problems, the network architecture, validation methods, and the reported quantitative performance criteria. The clinical relevance of the studies was then evaluated to identify limitations by considering the merits of convolutional neural networks and the remaining challenges that need to be solved to promote the clinical application and development of CNN algorithms. Finally, possible directions for future research are discussed for researchers in the biomedical and machine learning communities. A total of 83 studies were identified and reviewed. They differed in terms of the precise classification problem targeted and the strategies used to construct and train the chosen CNN. Consequently, the reported performance varied widely, with accuracies of 91.63–100% in differentiating meningiomas, gliomas, and pituitary tumors (26 articles) and of 60.0–99.46% in distinguishing low-grade from high-grade gliomas (13 articles). The review provides a survey of the state of the art in CNN-based deep learning methods for brain tumor classification. Many networks demonstrated good performance, and it is not evident that any specific methodological choice greatly outperforms the alternatives, especially given the inconsistencies in the reporting of validation methods, performance metrics, and training data encountered. Few studies have focused on clinical usability.

Management of brain tumors is based on clinical and radiological information with presumed grade dictating treatment. Hence, a non-invasive assessment of tumor grade is of paramount importance to choose the best treatment plan. Convolutional Neural Networks (CNNs) represent one of the effective Deep Learning (DL)-based techniques that have been used for brain tumor diagnosis. However, they are unable to handle input modifications effectively. Capsule neural networks (CapsNets) are a novel type of machine learning (ML) architecture that was recently developed to address the drawbacks of CNNs. CapsNets are resistant to rotations and affine translations, which is beneficial when processing medical imaging datasets. Moreover, Vision Transformers (ViT)-based solutions have been very recently proposed to address the issue of long-range dependency in CNNs. This survey provides a comprehensive overview of brain tumor classification and segmentation techniques, with a focus on ML-based, CNN-based, CapsNet-based, and ViT-based techniques. The survey highlights the fundamental contributions of recent studies and the performance of state-of-the-art techniques. Moreover, we present an in-depth discussion of crucial issues and open challenges. We also identify some key limitations and promising future research directions. We envisage that this survey shall serve as a good springboard for further study.

Machine learning models based on radiomic features allow us to obtain biomarkers that are capable of modeling the disease and that are able to support the clinical routine. Recent studies have shown that it is fundamental that the computed features are robust and reproducible. Although several initiatives to standardize the definition and extraction process of biomarkers are ongoing, there is a lack of comprehensive guidelines. Therefore, no standardized procedures are available for ROI selection, feature extraction, and processing, with the risk of undermining the effective use of radiomic models in clinical routine. In this study, we aim to assess the impact that the different segmentation methods and the quantization level (defined by means of the number of bins used in the feature-extraction phase) may have on the robustness of the radiomic features. In particular, the robustness of texture features extracted by PyRadiomics, and belonging to five categories—GLCM, GLRLM, GLSZM, GLDM, and NGTDM—was evaluated using the intra-class correlation coefficient (ICC) and mean differences between segmentation raters. In addition to the robustness of each single feature, an overall index for each feature category was quantified. The analysis showed that the level of quantization (i.e., the ‘bincount’ parameter) plays a key role in defining robust features: in fact, in our study focused on a dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) dataset of 111 breast masses, sets with cardinality varying between 34 and 43 robust features were obtained with ‘binCount’ values equal to 256 and 32, respectively. Moreover, both manual segmentation methods demonstrated good reliability and agreement, while automated segmentation achieved lower ICC values. Considering the dependence on the quantization level, taking into account only the intersection subset among all the values of ‘binCount’ could be the best selection strategy. Among radiomic feature categories, GLCM, GLRLM, and GLDM showed the best overall robustness with varying segmentation methods.

Several software tools for the simulation and analysis of biochemical reaction networks have been developed in the last decades; however, assessing and comparing their computational performance in executing the typical tasks of computational systems biology can be limited by the lack of a standardized benchmarking approach. To overcome these limitations, we propose here a novel tool, named SMGen, designed to automatically generate synthetic models of reaction networks that, by construction, are characterized by relevant features (e.g., system connectivity and reaction discreteness) and non-trivial emergent dynamics of real biochemical networks. The generation of synthetic models in SMGen is based on the definition of an undirected graph consisting of a single connected component that, generally, results in a computationally demanding task; to speed up the overall process, SMGen exploits a main–worker paradigm. SMGen is also provided with a user-friendly graphical user interface, which allows the user to easily set up all the parameters required to generate a set of synthetic models with any number of reactions and species. We analysed the computational performance of SMGen by generating batches of symmetric and asymmetric reaction-based models (RBMs) of increasing size, showing how a different number of reactions and/or species affects the generation time. Our results show that when the number of reactions is higher than the number of species, SMGen has to identify and correct a large number of errors during the creation process of the RBMs, a circumstance that increases the running time. Still, SMGen can generate synthetic models with hundreds of species and reactions in less than 7 s.

Unsupervised segmentation techniques, which do not require labeled data for training and can be more easily integrated into the clinical routine, represent a valid solution especially from a clinical feasibility perspective. Indeed, large-scale annotated datasets are not always available, undermining their immediate implementation and use in the clinic. Breast cancer is the most common cause of cancer death in women worldwide. In this study, breast lesion delineation in Dynamic Contrast Enhanced MRI (DCE-MRI) series was addressed by means of four popular unsupervised segmentation approaches: Split-and-Merge combined with Region Growing (SMRG), k-means, Fuzzy C-Means (FCM), and spatial FCM (sFCM). They represent well-established pattern recognition techniques that are still widely used in clinical research. Starting from the basic versions of these segmentation approaches, during our analysis, we identified the shortcomings of each of them, proposing improved versions, as well as developing ad hoc pre- and post-processing steps. The obtained experimental results, in terms of area-based—namely, Dice Index (DI), Jaccard Index (JI), Sensitivity, Specificity, False Positive Ratio (FPR), False Negative Ratio (FNR)—and distance-based metrics—Mean Absolute Distance (MAD), Maximum Distance (MaxD), Hausdorff Distance (HD)—encourage the use of unsupervised machine learning techniques in medical image segmentation. In particular, fuzzy clustering approaches (namely, FCM and sFCM) achieved the best performance. In fact, for area-based metrics, they obtained DI = 78.23% ± 6.50 (sFCM), JI = 65.90% ± 8.14 (sFCM), sensitivity = 77.84% ± 8.72 (FCM), specificity = 87.10% ± 8.24 (sFCM), FPR = 0.14 ± 0.12 (sFCM), and FNR = 0.22 ± 0.09 (sFCM). Concerning distance-based metrics, they obtained MAD = 1.37 ± 0.90 (sFCM), MaxD = 4.04 ± 2.87 (sFCM), and HD = 2.21 ± 0.43 (FCM). These experimental findings suggest that further research would be useful for advanced fuzzy logic techniques specifically tailored to medical image segmentation.

The Special Issue “Advanced Computational Methods for Oncological Image Analysis”, published for the Journal of Imaging, covered original research papers about state-of-the-art and novel algorithms and methodologies, as well as applications of computational methods for oncological image analysis, ranging from radiogenomics to deep learning [...]

Lung cancer (LC) is currently one of the main causes of cancer-related deaths worldwide. Low-dose computed tomography (LDCT) of the chest has been proven effective in secondary prevention (i.e., early detection) of LC by several trials. In this work, we investigated the potential impact of radiomics on indeterminate prevalent pulmonary nodule (PN) characterization and risk stratification in subjects undergoing LDCT-based LC screening. As a proof-of-concept for radiomic analyses, the first aim of our study was to assess whether indeterminate PNs could be automatically classified by an LDCT radiomic classifier as solid or sub-solid (first-level classification), and in particular for sub-solid lesions, as non-solid versus part-solid (second-level classification). The second aim of the study was to assess whether an LCDT radiomic classifier could automatically predict PN risk of malignancy, and thus optimize LDCT recall timing in screening programs. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC), accuracy, positive predictive value, negative predictive value, sensitivity, and specificity. The experimental results showed that an LDCT radiomic machine learning classifier can achieve excellent performance for characterization of screen-detected PNs (mean AUC of 0.89 ± 0.02 and 0.80 ± 0.18 on the blinded test dataset for the first-level and second-level classifiers, respectively), providing quantitative information to support clinical management. Our study showed that a radiomic classifier could be used to optimize LDCT recall for indeterminate PNs. According to the performance of such a classifier on the blinded test dataset, within the first 6 months, 46% of the malignant PNs and 38% of the benign ones were identified, improving early detection of LC by doubling the current detection rate of malignant nodules from 23% to 46% at a low cost of false positives. In conclusion, we showed the high potential of LDCT-based radiomics for improving the characterization and optimizing screening recall intervals of indeterminate PNs.

Biometric classification plays a key role in fingerprint characterization, especially in the identification process. In fact, reducing the number of comparisons in biometric recognition systems is essential when dealing with large-scale databases. The classification of fingerprints aims to achieve this target by splitting fingerprints into different categories. The general approach of fingerprint classification requires pre-processing techniques that are usually computationally expensive. Deep Learning is emerging as the leading field that has been successfully applied to many areas, such as image processing. This work shows the performance of pre-trained Convolutional Neural Networks (CNNs), tested on two fingerprint databases—namely, PolyU and NIST—and comparisons to other results presented in the literature in order to establish the type of classification that allows us to obtain the best performance in terms of precision and model efficiency, among approaches under examination, namely: AlexNet, GoogLeNet, and ResNet. We present the first study that extensively compares the most used CNN architectures by classifying the fingerprints into four, five, and eight classes. From the experimental results, the best performance was obtained in the classification of the PolyU database by all the tested CNN architectures due to the higher quality of its samples. To confirm the reliability of our study and the results obtained, a statistical analysis based on the McNemar test was performed.

Structural and metabolic imaging are fundamental for diagnosis, treatment and follow-up in oncology. Beyond the well-established diagnostic imaging applications, ultrasounds are currently emerging in the clinical practice as a noninvasive technology for therapy. Indeed, the sound waves can be used to increase the temperature inside the target solid tumors, leading to apoptosis or necrosis of neoplastic tissues. The Magnetic resonance-guided focused ultrasound surgery (MRgFUS) technology represents a valid application of this ultrasound property, mainly used in oncology and neurology. In this paper; patient safety during MRgFUS treatments was investigated by a series of experiments in a tissue-mimicking phantom and performing ex vivo skin samples, to promptly identify unwanted temperature rises. The acquired MR images, used to evaluate the temperature in the treated areas, were analyzed to compare classical proton resonance frequency (PRF) shift techniques and referenceless thermometry methods to accurately assess the temperature variations. We exploited radial basis function (RBF) neural networks for referenceless thermometry and compared the results against interferometric optical fiber measurements. The experimental measurements were obtained using a set of interferometric optical fibers aimed at quantifying temperature variations directly in the sonication areas. The temperature increases during the treatment were not accurately detected by MRI-based referenceless thermometry methods, and more sensitive measurement systems, such as optical fibers, would be required. In-depth studies about these aspects are needed to monitor temperature and improve safety during MRgFUS treatments.

Advances in microscopy imaging technologies have enabled the visualization of live-cell dynamic processes using time-lapse microscopy imaging. However, modern methods exhibit several limitations related to the training phases and to time constraints, hindering their application in the laboratory practice. In this work, we present a novel method, named Automated Cell Detection and Counting (ACDC), designed for activity detection of fluorescent labeled cell nuclei in time-lapse microscopy. ACDC overcomes the limitations of the literature methods, by first applying bilateral filtering on the original image to smooth the input cell images while preserving edge sharpness, and then by exploiting the watershed transform and morphological filtering. Moreover, ACDC represents a feasible solution for the laboratory practice, as it can leverage multi-core architectures in computer clusters to efficiently handle large-scale imaging datasets. Indeed, our Parent-Workers implementation of ACDC allows to obtain up to a 3.7× speed-up compared to the sequential counterpart. ACDC was tested on two distinct cell imaging datasets to assess its accuracy and effectiveness on images with different characteristics. We achieved an accurate cell-count and nuclei segmentation without relying on large-scale annotated datasets, a result confirmed by the average Dice Similarity Coefficients of 76.84 and 88.64 and the Pearson coefficients of 0.99 and 0.96, calculated against the manual cell counting, on the two tested datasets.

A clonogenic assay is a biological technique for calculating the Surviving Fraction (SF) that quantifies the anti-proliferative effect of treatments on cell cultures: this evaluation is often performed via manual counting of cell colony-forming units. Unfortunately, this procedure is error-prone and strongly affected by operator dependence. Besides, conventional assessment does not deal with the colony size, which is generally correlated with the delivered radiation dose or administered cytotoxic agent. Relying upon the direct proportional relationship between the Area Covered by Colony (ACC) and the colony count and size, along with the growth rate, we propose MF2C3, a novel computational method leveraging spatial Fuzzy C-Means clustering on multiple local features (i.e., entropy and standard deviation extracted from the input color images acquired by a general-purpose flat-bed scanner) for ACC-based SF quantification, by considering only the covering percentage. To evaluate the accuracy of the proposed fully automatic approach, we compared the SFs obtained by MF2C3 against the conventional counting procedure on four different cell lines. The achieved results revealed a high correlation with the ground-truth measurements based on colony counting, by outperforming our previously validated method using local thresholding on L*u*v* color well images. In conclusion, the proposed multi-feature approach, which inherently leverages the concept of symmetry in the pixel local distributions, might be reliably used in biological studies.

Prostate Cancer (PCa) is the most common oncological disease in Western men. Even though a growing effort has been carried out by the scientific community in recent years, accurate and reliable automated PCa detection methods on multiparametric Magnetic Resonance Imaging (mpMRI) are still a compelling issue. In this work, a Deep Neural Network architecture is developed for the task of classifying clinically significant PCa on non-contrast-enhanced MR images. In particular, we propose the use of Conditional Random Fields as a Recurrent Neural Network (CRF-RNN) to enhance the classification performance of XmasNet, a Convolutional Neural Network (CNN) architecture specifically tailored to the PROSTATEx17 Challenge. The devised approach builds a hybrid end-to-end trainable network, CRF-XmasNet, composed of an initial CNN component performing feature extraction and a CRF-based probabilistic graphical model component for structured prediction, without the need for two separate training procedures. Experimental results show the suitability of this method in terms of classification accuracy and training time, even though the high-variability of the observed results must be reduced before transferring the resulting architecture to a clinical environment. Interestingly, the use of CRFs as a separate postprocessing method achieves significantly lower performance with respect to the proposed hybrid end-to-end approach. The proposed hybrid end-to-end CRF-RNN approach yields excellent peak performance for all the CNN architectures taken into account, but it shows a high-variability, thus requiring future investigation on the integration of CRFs into a CNN.