SPIE Conference Part 3: Region of Interest Detection May Improve WSI Classification Performance of Patch-based Models

by Christos Evangelou, MSc, PhD – Medical Writer and Editor

SAN DIEGO, California – In a new study, researchers at Bilkent University in Turkey investigated the effect of region of interest (ROI) detection on the classification performance of patch-based models when applied to whole slide images (WSIs). The researchers compared the WSI classification performance of patch-based models when the patches were extracted from the entire image and when they were sampled only within the ROIs. The study showed that using the predicted ROIs for patch sampling significantly improved WSI classification. The study findings were presented at SPIE Medical Imaging 2023, which took place in San Diego on February 19–23.

WSI image classification typically involves the use of fixed-size patches to make slide-level predictions. A key limitation of this approach when used for diagnostic purposes is that not all batches contain equally important histopathological information.

“Another issue with fixed-size patches is that patches are treated individually, and contextual information is often not used in this process,” said Dr Selim Aksoy, professor and senior investigator at Bilkent University. “To overcome these issues, we studied whether we can benefit from ROI detection for WSI classification,” he added.

The first step in the proposed WSI classification pipeline is to obtain binary segmentation of the ROIs for each WSI. To achieve this, the team used both single- and multi-resolution models. “The goal of this step is to produce a slide-wide pixel-level segmentation map for detecting ROIs,” Dr Aksoy explained.

Subsequently, patches sampled from the output of the ROI detection algorithm are used as input for the WSI classifier. Finally, the WSI classifier reaches a slide-level prediction (i.e., benign or malignant) by comparing the average probability of the patches to a threshold.

The researchers found that single-resolution models (U-Net) and multi-resolution models (pyramidal and HookNet) provided similar ROI detection in terms of precision, recall, and F1 scores. “We were expecting that multi-resolution models would have performed better because that’s what the pathologists are really looking at. But the issue was that these more complex models had some convergence issues,” Dr Aksoy noted.

The study also showed that, compared to sampling patches from the entire WSI, model training and testing using patches only from detected ROIs improved the WSI classification performance of patch-based models in terms of accuracy, precision, recall, and F1 scores. The validation dataset included 98 WSIs of H&E-stained breast tissues from 81 patients.

“We believe that further improvements are possible with additional refinement of ROI masks, better parameter tuning, and longer training epochs,” Dr Aksoy said.

The study was supported in part by the Outstanding Young Scientists Program (GEBIP) of the Turkish Academy of Sciences.