Encoder-Decoder Variant Analysis for Semantic Segmentation of Gastrointestinal Tract Using UW-Madison Dataset

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Gastrointestinal (GI) tract cancer poses a significant global health challenge.
• Accurate segmentation of GI organs is critical for effective radiation therapy planning.
• Current methods require precise delineation of the stomach, small bowel, and large bowel to minimize damage to healthy tissues.

Purpose of the Study:

• To explore and identify optimal encoder-decoder architectures for segmenting GI organs in MRI images.
• To evaluate the performance of various deep learning models for precise organ segmentation.
• To enhance the accuracy of radiation targeting in GI cancer treatment through improved image segmentation.

Main Methods:

• Utilized the UW-Madison GI tract dataset comprising 38,496 MRI scans.
• Tested multiple encoder networks (ResNet50, EfficientNetB1, MobileNetV2, ResNext50, Timm_Gernet_S).
• Paired encoders with decoders including UNet, FPN, PSPNet, PAN, and DeepLab V3+ for segmentation.

Main Results:

• The combination of ResNet50 encoder and DeepLab V3+ decoder demonstrated superior performance.
• Achieved a Dice coefficient of 0.9082 and a Jaccard index (IoU) of 0.8796.
• The model yielded a low model loss of 0.117, indicating high segmentation accuracy.

Conclusions:

• The ResNet50-DeepLab V3+ model offers a highly effective solution for GI organ segmentation in MRI.
• This approach has the potential to significantly improve radiation therapy planning and delivery for GI cancer patients.
• Findings support the integration of advanced AI in biomedical image analysis for clinical applications.