Exploration de la pr´ediction de la r´ecidive du cancer du sein `a partir d’images histopathologiques en utilisant l’intelligence artificielle

Abstract

This study presents a comparative analysis of multiple classification models for predicting breast cancer recurrence from histopathological features, based on a deep learning pipeline developed in prior work. The referenced pipeline employs a CNN-based discriminator to score patches extracted from whole-slide images (WSIs), selects the top most discriminative patches per slide using the Discrimination Score (DS), extracts deep feature vectors, and classifies recurrence risk at the slide level using a Multiple Instance Learning (MIL) framework. Our contribution lies in the comprehensive evaluation and comparison of several classification algorithms including Decision Trees, Bagging, Gaussian Naive Bayes, and K-Nearest Neighbors on the same dataset and features produced by the existing pipeline. The goal was to investigate their effectiveness, generalizability, and clinical relevance for breast cancer recurrence prediction. The MIL classifier, forming part of the original pipeline, achieved a slide-level AUC of 0.82 and an accuracy of 78.5%, outperforming all traditional models. Among the latter, Bagging demonstrated relatively strong results (AUC of 0.84), yet suffered from a high false discovery rate (77.8%). Decision Trees showed decent validation performance (AUC up to 0.76) but overfitted the training data. Gaussian Naive Bayes achieved high precision (92.3% PPV) for a single class but failed to generalize across classes. KNN reached an AUC of 0.85 during validation, which dropped significantly to 0.72 in testing, indicating poor robustness. This comparative study reinforces the advantages of MIL-based models for handling weakly labeled medical data and highlights the limitations of classical approaches in complex, high-dimensional imaging tasks. Our findings emphasize the importance of both intelligent patch selection and advanced learning frameworks in improving diagnostic reliability and decision support in digital pathology.