Overview / Usage

We propose a fully automated pipeline based on deep learning, for signet-ring cell detection. Signet-ring cells are large vacuole cells that are predominantly found in carcinomas or gastric cancer. Early-stage detection of gastric cancer helps in resection of these signet-ring cells endoscopically. We make use of a patch-based approach for detection, which is necessary for extending our work on whole slide images. we make use of the Mask R-CNN approach, which involves multi-resolution feature extraction block, region proposal block, followed by classifier head. We make use of ResNet-X backbone for feature extraction, and collect output from various intermediate layers which serves as multi-resolution features, encoding the information in various scales. These multi-resolution features are then used by region proposal network which regresses for bounding box and confidence value estimation. All the proposed regions are used in gradient update via the classifier and regressor loss. We made use of multiple augmentation methods to increase the generalizability of the model. The model was also trained on negative images using a hard-mining strategy. During inference we extract overlapping patches and estimate bounding boxes and classify them accordingly, followed by radius based averaging to obtain one fine detection for one signet-ring cell. The model was trained on 80% data and was validated on the remaining 20% with a recall score of 0.6125 and zero normal false positives.

​

Methodology / Approach

MaskRCNN based detection algorithm

Technologies Used

pytorch, numpy, xml, pandas