Despite recent advances in digital pathology, the development of a robust deep learning model for identifying tumor regions on digitalized slide images has been plagued by problems related to the patch-based approach attributable to the narrow and limited field-of-views. In this study, we explored the effect of the fused lasso technique on diagnostic performance using gastric cancer pathology images. Our goal is to develop a speedy tumor segmentation scheme that overcomes performance degradation arising from the limitations of the patch-based approach. We used 27 specimens with partial annotation from tissue-microarrays (TMAs) with gastric cancer from the Asan Medical Center. Our segmentation adjustment method was tested on multiple renowned deep learning model architectures to detect malignancy on tissue slides. In the fused lasso formulation, fusion and lasso penalty were adopted to penalize variables with L1-norms based on both the patchwise predictions and their pairwise differences between adjoining patches. To obtain well-structured boundaries and discard irrelevant observations, the best performing hyperparameters controlling the degree to be penalized were adaptively determined via cross-validation. We demonstrated that adoption of fused lasso penalties can produce segmentation close to that generated by experts without a significant increase in computing time. The average test performance in terms of IOU, dice, accuracy, precision, recall, and AUROC increased by a degree of 3.74%p, 3.27%p, 0.93%p, 5.32%p, 3.20%p, and 1.58%p, respectively. This study provides a use case of the application of fused lasso modeling for precisely identifying tumor area, which can alleviate the bottleneck in gastric cancer research.

Citation Format: Jiwon Jung, Jin Roh, Chan-Sik Park. Fused LASSO application for gastric cancer image segmentation [abstract]. In: Proceedings of the AACR Virtual Special Conference on Artificial Intelligence, Diagnosis, and Imaging; 2021 Jan 13-14. Philadelphia (PA): AACR; Clin Cancer Res 2021;27(5_Suppl):Abstract nr PO-079.