SIR 2024
Imaging
Hossam A. Zaki, BS
Medical Student
Warren Alpert Medical School of Brown University
Financial relationships: Full list of relationships is listed on the CME information page.
Scott Collins, RT(R)(CT)
Imaging Clinical Specialist
Rhode Island Hospital
Disclosure information not submitted.
Harrison Bai, MD, MS
Director, Radiology AI Lab
Johns Hopkins University
Disclosure information not submitted.
Aaron W. Maxwell, MD
Professor
Alpert Medical School, Brown University
Disclosure information not submitted.
Segmentation of liver cancer from computed tomography (CT) scans is often the first step in the treatment and is often done manually. This project tests the efficacy of two novel machine-learning models in the task of liver tumor segmentation.
Materials and methods:
This HIPAA-compliant study was performed with a waiver for informed consent following institutional review board approval. Between December 11, 2018, and October 11, 2022, adult patients who underwent image-guided thermal ablation of liver tumors were retrospectively identified. A pretrained U-Net to segment liver zones, with a DICE score of 0.946, was applied to a dataset of pre-procedure CT scans {1}. Following this, a U-shaped encoder-decoder Transformer architecture (UNETR) as well as MedNeXt, an additional transformer architecture claiming to have superior performance over UNETR and built for smaller datasets, was trained to segment liver tumors. Both models have yet to be tested for liver segmentation.
Results:
50 CT scans were extracted. The dataset was split into 80% training and 20% testing. Our pipeline, despite its limited data, achieved state-of-the-art performance in this task. UNETR and MedNeXt achieved an average DICE of 0.92 and 0.88, respectively. This indicates a percent overlap between the predicted segmentation and the ground truth. UNETR and MedNeXt achieved a MeanIoU of 0.91, and .88, respectively. This calculates the average ratio of the area of overlap to the area of union between the predicted and ground truth segments. Upon qualitative review, the pipeline was able to capture the fine-grained details, despite the similarity in the liver field and tumor.
Conclusion:
Our proposed pipeline shows that transformer-based models are able to accurately segment liver tumors after extracting liver fields, achieving a state-of-the-art performance with limited data. Tumor segmentation is the first step in many tasks, both clinical and technological. It is important in treatment planning, generating impressions, determining prognosis, and being involved in multiple radiomics pipelines. An automatic segmentation method that can achieve high performance with limited data, as is the case with a lot of imaging data, is of substantial clinical need.