MSTT-199: MRI Dataset for Musculoskeletal Soft Tissue Tumor Segmentation

TL;DR

The collection of an MR imaging dataset of 199 musculoskeletal soft tissue tumors from 199 patients was described and the model predictions were analyzed and found that its performance suffered on fibrous and vascular tumors due to their diverse anatomical location, size, and intensity heterogeneity.

Abstract

Accurate musculoskeletal soft tissue tumor segmentation is vital for assessing tumor size, location, diagnosis, and response to treatment, thereby influencing patient outcomes. However, segmentation of these tumors requires clinical expertise, and an automated segmentation model would save valuable time for both clinician and patient. Training an automatic model requires a large dataset of annotated images. In this work, we describe the collection of an MR imaging dataset of 199 musculoskeletal soft tissue tumors from 199 patients. We trained segmentation models on this dataset and then benchmarked them on a publicly available dataset. Our model achieved the state-of-the-art dice score of 0.79 out of the box without any fine tuning, which shows the diversity and utility of our curated dataset. We analyzed the model predictions and found that its performance suffered on fibrous and vascular tumors due to their diverse anatomical location, size, and intensity heterogeneity. The code and models are available in the following github repository, https://github.com/Reasat/mstt

Figures (17)

• Figure 1: The Label Studio annotation setup used for this project. The annotators can choose a brush size and annotate the tumor at different granularity.
• Figure 2: Example of different tissue types present in the MSTT-199 corpus. The Fat tumor is brighter on T1 while the rest of the tissue types are brighter on T2.
• Figure 3: Tumor volume distribution. Fat tumors were on average larger than those for the other subtypes, while the other four subtypes were more similar.
• Figure 4: Average tumor intensity distribution in T1 and T2 image. Myxoid tumors tended to have a lower intensity on average on both T1 and T2, while the other tumors were more similar in intensity.
• Figure 5: The segmentation model architectures. The input to both the segmentation models is a multi-modal MRI image (T1 and T2). The 2.5-D slices coming from each modality are stacked to create a six-channel input. \ref{['fig:unet']}) The U-Net segmentation model has a U-shaped structure due to skip connections going from the encoders to the decoders. \ref{['fig:sam']}) The SAM model has an additional prompt encoder. For automatic segmentation, the prompt encoder receives a bounding box drawn over the full image.