Deep learning-based interpretable prediction of recurrence of diffuse large B-cell lymphoma

Area of Science:

• Computational pathology
• Artificial intelligence in oncology
• Cancer imaging analysis

Background:

• Diffuse large B-cell lymphoma (DLBCL) is aggressive, with up to 50% of patients experiencing recurrence after chemotherapy.
• Predicting recurrence early is crucial for initiating alternative treatment strategies.
• Current deep learning models for cancer recurrence prediction lack interpretability.

Purpose of the Study:

• To develop an interpretable deep learning pipeline for predicting DLBCL recurrence from histological images.
• To identify key image-based features associated with DLBCL recurrence.
• To enhance understanding of the biological basis of DLBCL heterogeneity and treatment outcomes.

Main Methods:

• Developed a deep learning pipeline using histological images from a public DLBCL cohort.
• Employed attention-based classification to identify relevant image regions.
• Segmented nuclei in high-relevance areas and analyzed morphological features for recurrence prediction.

Main Results:

• Achieved an f1 score of 0.88, demonstrating high accuracy in distinguishing between patients who recurred and those who did not.
• Identified large and irregularly shaped tumor cell nuclei as the most predictive features for recurrence.
• Highlighted the utility of attention mechanisms in pinpointing diagnostically relevant regions within histological images.

Conclusions:

• Histological images hold significant value for predicting treatment outcomes in DLBCL.
• The study enhances the understanding of complex biological processes driving aggressive cancers like DLBCL.
• Interpretable deep learning models can provide insights into cancer recurrence, potentially leading to novel therapeutic approaches.