Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Weakly-Supervised Detection of Bone Lesions in CT

Tao Sheng, Tejas Sudharshan Mathai, Alexander Shieh, Ronald M. Summers

TL;DR

The paper tackles automated detection of bone metastases in CT using weak supervision by converting radiologist RECIST measurements into proxy 3D masks through GrabCut and extending them to partial 3D annotations. These masks, merged with skeletal/body context from TotalSegmentator, train a 3D full-resolution nnUNet that delivers high precision (≈96.7%) but modest recall (≈47.3%) after ground-truth review. The study demonstrates that learning from incomplete annotations can localize bone lesions and support tracking across CT volumes, while highlighting the need for more comprehensive voxel-level ground-truth data to boost sensitivity. It underscores the balance between precision and practicality in CAD systems and outlines future work to scale up annotation and improve lesion detection coverage.

Abstract

The skeletal region is one of the common sites of metastatic spread of cancer in the breast and prostate. CT is routinely used to measure the size of lesions in the bones. However, they can be difficult to spot due to the wide variations in their sizes, shapes, and appearances. Precise localization of such lesions would enable reliable tracking of interval changes (growth, shrinkage, or unchanged status). To that end, an automated technique to detect bone lesions is highly desirable. In this pilot work, we developed a pipeline to detect bone lesions (lytic, blastic, and mixed) in CT volumes via a proxy segmentation task. First, we used the bone lesions that were prospectively marked by radiologists in a few 2D slices of CT volumes and converted them into weak 3D segmentation masks. Then, we trained a 3D full-resolution nnUNet model using these weak 3D annotations to segment the lesions and thereby detected them. Our automated method detected bone lesions in CT with a precision of 96.7% and recall of 47.3% despite the use of incomplete and partial training data. To the best of our knowledge, we are the first to attempt the direct detection of bone lesions in CT via a proxy segmentation task.

Weakly-Supervised Detection of Bone Lesions in CT

TL;DR

The paper tackles automated detection of bone metastases in CT using weak supervision by converting radiologist RECIST measurements into proxy 3D masks through GrabCut and extending them to partial 3D annotations. These masks, merged with skeletal/body context from TotalSegmentator, train a 3D full-resolution nnUNet that delivers high precision (≈96.7%) but modest recall (≈47.3%) after ground-truth review. The study demonstrates that learning from incomplete annotations can localize bone lesions and support tracking across CT volumes, while highlighting the need for more comprehensive voxel-level ground-truth data to boost sensitivity. It underscores the balance between precision and practicality in CAD systems and outlines future work to scale up annotation and improve lesion detection coverage.

Abstract

The skeletal region is one of the common sites of metastatic spread of cancer in the breast and prostate. CT is routinely used to measure the size of lesions in the bones. However, they can be difficult to spot due to the wide variations in their sizes, shapes, and appearances. Precise localization of such lesions would enable reliable tracking of interval changes (growth, shrinkage, or unchanged status). To that end, an automated technique to detect bone lesions is highly desirable. In this pilot work, we developed a pipeline to detect bone lesions (lytic, blastic, and mixed) in CT volumes via a proxy segmentation task. First, we used the bone lesions that were prospectively marked by radiologists in a few 2D slices of CT volumes and converted them into weak 3D segmentation masks. Then, we trained a 3D full-resolution nnUNet model using these weak 3D annotations to segment the lesions and thereby detected them. Our automated method detected bone lesions in CT with a precision of 96.7% and recall of 47.3% despite the use of incomplete and partial training data. To the best of our knowledge, we are the first to attempt the direct detection of bone lesions in CT via a proxy segmentation task.
Paper Structure (7 sections, 2 figures, 1 table)

This paper contains 7 sections, 2 figures, 1 table.

Table of Contents

  1. INTRODUCTION
  2. METHODS
  3. Dataset
  4. Generation of Ground Truth Data
  5. Detection of Bone Lesions using nnUNet
  6. Experiments and Results
  7. DISCUSSION AND CONCLUSION

Figures (2)

  • Figure 1: Framework for the detection of bone lesions via a proxy segmentation task. First, the prospective RECIST measurements (magenta) were used to obtain the lesion mask (blue) via GrabCut. Then, the enclosing bounding box of the lesion mask was extended to create weak 3D masks. Next, the body region (red) and the skeletal region (green) from TotalSegmentator was also merged with the weak 3D masks. Finally, a 3D full-resolution nnUNet model was trained to predict the corresponding regions in the CT volume where the bone lesions were observed. At test time, the model received a 3D CT volume and detected bone lesions in it.
  • Figure 2: Qualitative results of bone lesion detection. Each column shows a cropped slice with the top row displaying ground-truth (green) and the bottom row showing nnUNet predictions (yellow). (a) shows a TP osteolytic lesion; (b) shows the prediction of a real osteoblastic lesion that was not prospectively marked as ground-truth (before review), and after GT review, it was correctly considered as ground-truth; (c) shows an osteolytic lesion not annotated in ground-truth before review and correctly marked afterwards. A FN adjacent to it is also shown. (d) shows a rare FP on the cortex of a rib.