255 Predicting patient-level extranodal extension using pre-treatment computed tomography imaging.

Extranodal extension (ENE) is strongly associated with worse outcomes in patients with head and neck squamous cell carcinoma (HNSCC) and is included in staging systems for HNSCC2,3. While histopathological determination remains the gold standard, pathologic confirmation is not always available for every patient or may not be available at the time when evidence of ENE may alter the treatment plan. This study aims to develop an automated method of pre-treatment radiographic imaging ENE (iENE) detection without human intervention to increase the chances of clinical adoption and improve risk stratification of patients with ENE. In a previous study by Kann et al, a deep learning model was used to predict the presence of iENE in pre-treatment computed tomography (CT) scans of lymph nodes (LNs) cropped along their borders1. The model trained on 2673 LNs from 270 patients achieved a notable AUC score of 0.91 on the held-out test set, and 0.84 when tested on multi-institutional data. This performance gap may be attributed to the strong correlation between ENE status and LN diameter, as revealed by their regression analysis of clinical variables. Additionally, deploying their model in clinical settings necessitates substantial human involvement, as it requires contouring single or multiple suspicious LNs for inference. To address these issues, we introduce a deep learning model designed to predict patient-level ENE without LN segmentations. Not relying on manual segmentation or cropping along the LNs borders improves potential adoption and serves as a regularization technique to reduce feature correlation with LN volume. We utilized a retrospective dataset comprising 922 oropharyngeal cancer patients and their radiotherapy (RT) treatment planning CT scans, each annotated by a radiologist for the presence of iENE. Our model analyzes 256x256x128 voxels centered around the larynx, encompassing LNs across all neck levels. Due to the input's considerable size, conventional 3D convolutional neural networks proved inefficient due to GPU memory limitations. To circumvent this, we leveraged ACS convolutions (ACSConv), which adapt 2D convolutions for 3D volumetric images, allowing us to capitalize on ImageNet pre-trained weights for more robust feature learning and quicker convergence. Our dataset was divided into 759 cases for training and 163 for held-out testing, with the date of diagnosis as the separation point, simulating a pseudo-prospective in-silico trial. The model was trained across four cross-validation folds on the training set, and the results from these four models were ensembled during inference. Our ACSConv models, without pre-training, achieved an area under the receiver operating characteristic curve (AUROC) of 0.80-0.83 and an area under the precision-recall curve (AUPRC) of 0.47-0.61 on the held-out test set, whereas our ImageNet pre-trained ACSConv model improved with an AUROC of 0.87-0.91 and an AUPRC of 0.52-0.69. This highlights the benefit of using ImageNet pre-trained weights, which encapsulate robust features learned from millions of standard images. However, it's crucial to recognize that these performance metrics solely gauge the model's alignment with iENE, not histopathology. To demonstrate the model's risk stratification capabilities, we examined overall survival using Kaplan-Meier curves. The radiologist's iENE determination did not yield a statistically significant difference (p=0.01), whereas the model's predictions resulted in significant differences (p<0.005). The increased prognostic power may be attributed to inherent limitations of radiological determination compared to histopathology. Alternatively, the model may be less sensitive and highlight severe cases, increasing its prognostic power. External validation on unseen cohorts will further elucidate the model's prognostic capabilities and limitations. [Display omitted] Additionally, we assessed the model's predictions against the volume of the largest LN gross tumor volume (GTVn) and found a moderate Spearman's correlation coefficient of 0.531. This suggests that while the model does consider GTVn volume, it certainly incorporates information beyond LN borders, potentially shedding light on issues with GTVn determination. [Display omitted] To account for potential variability in larynx segmentation models, we explored the impact of test-time augmentations on model inference. This simulated variations in larynx segmentation and its effect on iENE model performance. Notably, we observed no significant differences in prediction standard deviations between patients with or without iENE. An interesting trend emerged, showing that prediction standard deviations increased as the model's confidence decreased (closer to 0.5). This underscores the necessity for a deeper investigation into deep learning model behaviors to enhance safety and increase the likelihood of clinical adoption. [Display omitted] Our study presents a novel deep learning model for the automated detection of pre-treatment iENE. With improved performance and prognostic capabilities, this model holds potential for enhancing risk stratification in HNSCC patients and merits further validation on external datasets and in clinical practice. [ABSTRACT FROM AUTHOR]