Search

Abstract Deep neural networks (DNNs) show promise in image-based medical diagnosis, but cannot be fully trusted since they can fail for reasons unrelated to underlying pathology. Humans are less likely to make such superficial mistakes, since they use features that are grounded on medical science. It is therefore important to know whether DNNs use different features than humans. Towards this end, we propose a framework for comparing human and machine perception in medical diagnosis. We frame the comparison in terms of perturbation robustness, and mitigate Simpson’s paradox by performing a subgroup analysis. The framework is demonstrated with a case study in breast cancer screening, where we separately analyze microcalcifications and soft tissue lesions. While it is inconclusive whether humans and DNNs use different features to detect microcalcifications, we find that for soft tissue lesions, DNNs rely on high frequency components ignored by radiologists. Moreover, these features are located outside of the region of the images found most suspicious by radiologists. This difference between humans and machines was only visible through subgroup analysis, which highlights the importance of incorporating medical domain knowledge into the comparison.

Ultrasound is an important imaging modality for the detection and characterization of breast cancer, but it has been noted to have high false-positive rates. Here, the authors present an artificial intelligence system that achieves radiologist-level accuracy in identifying breast cancer in ultrasound imaging.

Abstract The purpose of this work was to develop a novel method to disentangle the intra- and extracellular components of the total sodium concentration (TSC) in breast cancer from a combination of proton ( $$^{1}$$ 1 H) and sodium ( $$^{23}\hbox {Na}$$ 23 Na ) magnetic resonance imaging (MRI) measurements. To do so, TSC is expressed as function of the intracellular sodium concentration ( $$\hbox {C}_{\text {IC}}$$ C IC ), extracellular volume fraction (ECV) and the water fraction (WF) based on a three-compartment model of the tissue. TSC is measured from $$^{23}\hbox {Na}$$ 23 Na MRI, ECV is calculated from baseline and post-contrast $$^{1}$$ 1 H $$\hbox {T}_{{1}}$$ T 1 maps, while WF is measured with a $$^{1}$$ 1 H chemical shift technique. $$\hbox {C}_{\text {IC}}$$ C IC is then extrapolated from the model. Proof-of-concept was demonstrated in three healthy subjects and two patients with triple negative breast cancer. In both patients, TSC was two to threefold higher in the tumor than in normal tissue. This alteration mainly resulted from increased $$\hbox {C}_{\text {IC}}$$ C IC ( $$\sim$$ ∼ 30 mM), which was $$\sim$$ ∼ 130% greater than in healthy conditions (10–15 mM) while the ECV was within the expected range of physiological values (0.2–0.25). Multinuclear MRI shows promise for disentangling $$\hbox {C}_{\text {IC}}$$ C IC and ECV by taking advantage of complementary $$^{1}$$ 1 H and $$^{23}\hbox {Na}$$ 23 Na measurements.

Objective: To examine the prognostic capabilities of intravoxel incoherent motion (IVIM) metrics and their ability to predict response to neoadjuvant treatment (NAT). Additionally, to observe changes in IVIM metrics between pre- and post-treatment MRI. Methods: This IRB-approved, HIPAA-compliant retrospective study observed 31 breast cancer patients (32 lesions). Patients underwent standard bilateral breast MRI along with diffusion-weighted imaging before and after NAT. Six patients underwent an additional IVIM-MRI scan 12â14 weeks after initial scan and 2 cycles of treatment. In addition to apparent diffusion coefficients (ADC) from monoexponential decay, IVIM mean values (tissue diffusivity Dt, perfusion fraction fp, and pseudodiffusivity Dp) and histogram metrics were derived using a biexponential model. An additional filter identified voxels of highly vascular tumor tissue (VTT), excluding necrotic or normal tissue. Clinical data include histology of biopsy and clinical response to treatment through RECIST assessment. Comparisons of treatment response were made using Wilcoxon rank-sum tests. Results: Average, kurtosis, and skewness of pseudodiffusion Dp significantly differentiated RECIST responders from nonresponders. ADC and Dt values generally increased (â¼70%) and VTT% values generally decreased (â¼20%) post-treatment. Conclusion: Dp metrics showed prognostic capabilities; slow and heterogeneous pseudodiffusion offer poor prognosis. Baseline ADC/Dt parameters were not significant predictors of response. This work suggests that IVIM mean values and heterogeneity metrics may have prognostic value in the setting of breast cancer NAT. Keywords: Breast cancer, Diffusion weighted MRI, Intravoxel incoherent motion, Neoadjuvant treatment, Response evaluation criteria in solid tumors

Purpose. In November 2009, the U.S. Preventative Service Task Force (USPSTF) revised their breast cancer screening guidelines. We evaluated the pattern of screening subsequent to the altered guidelines in a cohort of women. Methods. Our database was queried for the following variables: age, race, method of diagnosis, mass palpability, screening frequency, histology, and stage. Statistical analyses were performed using Pearson’s chi-square and Fisher’s exact tests. Results. 1112 women were diagnosed with breast cancer from January 2010 to 2012. The median age at diagnosis was 60 years. Most cancers were detected on mammography (61%). The majority of patients had invasive ductal carcinoma (59%), stage 0 (23%), and stage 1 (50%) cancers. The frequency of screening did not change significantly over time (P=0.30). However, nonregular screeners had an increased risk of being diagnosed with later stage breast cancer (P