Synthetic DOmain-Targeted Augmentation (S-DOTA) Improves Model Generalization in Digital Pathology

• DOI: 10.48550/arxiv.2305.02401
• Publication Date: 2023-01-01
• Authors: Sai Chowdary Gull...; Yibo Zhang; Nitin Kumar Mittal; Deeksha Kartik; Sandhya Srinivasan; Kevin C. Rose; Daniel Shenker; Dinkar Juyal; Harshith Padigela; Raymond Biju; Victor Minden; Chirag Maheshwari; Marc Thibault; Zvi Goldstein; Luke A. Novak; Nidhi Chandra; Justin Lee; Aaditya Prakash; Chintan Shah; John F. Abel; Darren Fahy; Amaro Taylor‐Weiner; Anand Sampat
• Keywords: FOS: Computer and information sciences; Computer Science - Machine Learning; Computer Vision and Pattern Recognition (cs.CV); Computer Science - Computer Vision and Pattern Recognition; Machine Learning (cs.LG)

Machine learning algorithms have the potential to improve patient outcomes in digital pathology. However, generalization of these tools is currently limited by sensitivity variations tissue preparation, staining procedures and scanning equipment that lead domain shift digitized slides. To overcome this limitation model generalization, we studied effectiveness two Synthetic DOmain-Targeted Augmentation (S-DOTA) methods, namely CycleGAN-enabled Scanner Transform (ST) targeted Stain Vector (SVA), compared them against International Color Consortium (ICC) profile-based color calibration (ICC Cal) method a baseline using traditional brightness, noise augmentations. We evaluated ability techniques various tasks settings: four models, types (tissue segmentation cell classification), loss functions, six labs, scanners, three indications (hepatocellular carcinoma (HCC), nonalcoholic steatohepatitis (NASH), prostate adenocarcinoma). methods based on macro-averaged F1 scores in-distribution (ID) out-of-distribution (OOD) test sets across multiple domains, found S-DOTA (i.e., ST SVA) led significant improvements over ICC Cal OOD data while maintaining comparable performance ID data. Thus, demonstrate may help address due real world applications.