MYC Rearrangement Prediction From LYSA Whole Slide Images in Large B-Cell Lymphoma: A Multicentric Validation of...

Area of Science:

• Computational pathology
• Digital pathology
• Oncology

Background:

• Large B-cell lymphoma (LBCL) is a diverse cancer with MYC gene rearrangement (MYC-R) indicating a poor prognosis.
• Current MYC-R detection via fluorescence in situ hybridization (FISH) is costly, slow, and not universally accessible.

Purpose of the Study:

• To develop an automated, interpretable deep learning algorithm for MYC-R detection in LBCL using hematoxylin-and-eosin (H&E) stained whole slide images.
• To reduce reliance on FISH testing and enhance diagnostic efficiency for pathologists.

Main Methods:

• An interpretable deep learning model was created, incorporating self-supervised learning techniques.
• The model was trained and validated on four independent multicentric cohorts comprising 1247 LBCL patients.
• Extensive comparisons were performed across 7 feature extractors and 6 multiple instance learning models.

Main Results:

• The best deep learning model achieved an average AUC of 81.9% on cross-validation and AUCs from 62.2% to 74.5% on unseen cohorts.
• The model demonstrated potential as a prescreening tool, potentially eliminating the need for FISH in 35% of cases with a 0% false-negative rate.

Conclusions:

• Automated MYC-R detection on H&E slides using deep learning is feasible for routine pathology practice.
• This approach offers a cost-effective and efficient alternative to traditional molecular testing for LBCL.
• The developed algorithm could serve as a valuable medical device for improving LBCL diagnosis and patient management.