Crop-Guided Neural Network Segmentation of High-Resolution Skin Lesion Images

Medical images are often exceedingly large in width and height, limiting the maximum batch size when training convolutional neural networks and requiring models with a large number of parameters. Typically, images are uniformly downsampled, leading to losing fine-detailed information. Instead of uniformly downsampling images, we introduce a two-stage end-to-end segmentation network utilizing image crops to reduce network input size. Initially, a uniformly downscaled image is first segmented with a rough segmentation module, and the rough segmentation is used as a saliency map to crop the original high-resolution image to a region of interest. This crop is then re-segmented with a fine segmentation module. Our method’s effectiveness is demonstrated in segmenting lesion boundaries in clinical images across two datasets. We establish that this technique maintains comparable segmentation quality to a baseline model while reducing the network input size. Furthermore, our approach enhances the robustness of segmentation outcomes with smaller input sizes, outperforming uniformly downscaled images and baseline models. This improvement is consistent in both in-sample and out-of-sample evaluations.