Issues with Baseline UNet for Tissue Segmentation Using Provided GitHub Code

[Younger330]

I've been using the baseline UNet model from your GitHub repository for segmenting tissue regions. However, I'm encountering issues where the prediction heatmaps for Cancer Area and Unknown almost exclusively 0, while the heatmap for Tumor Area tends toward 1.

I've implemented cross-entropy and Dice loss, applying softmax across all three channels. Additionally, I've tried both resizing and not resizing the input images, yet neither approach yields correct results. Could you provide guidance or potential adjustments to address this issue?