We study lesion detection, segmentation, classification, and quantitative diagnostic support using endoscopic images, CT images, cellular images, and pathological images.
The laboratory combines deep learning with optics-, geometry-, and physics-based image understanding so that medical AI can not only detect lesions, but also estimate shape, depth, and size.
3D reconstruction of endoscopic polyps from monocular or sparse-view images is one of our characteristic research topics.
Why It Is Difficult
Medical images are affected by institution-specific acquisition protocols, device differences, patient variability, and lesion diversity. Small datasets and class imbalance are common.
Endoscopy further involves non-rigid deformation, specular highlights, illumination changes, and camera motion, making shape reconstruction and absolute-scale estimation from a single view intrinsically difficult.
Practical systems must also satisfy clinical constraints such as acceptable false positives, measurement stability, explainability, and workflow compatibility.
Approach
We connect data-driven deep learning with optical, geometric, and physical constraints, treating detection and 3D reconstruction as a connected problem.
1. Lesion detection, classification, and segmentation
We use transfer learning, semi-supervised learning, generative models, and data augmentation to build robust methods for endoscopic, CT, cellular, and pathological images.
2. 3D shape reconstruction and quantitative measurement
We estimate relative depth, 3D shape, and absolute size from monocular or sparse-view images using cues such as vascular structure, shading, illumination, and projection models.
3. Domain shift handling
We design medical image AI that remains stable under institution, device, and imaging-condition shifts through domain adaptation and data augmentation.
4. Integration with clinical workflows
We evaluate whether the methods reduce missed lesions, improve measurement reproducibility, and shorten diagnostic workflows.
Evaluation
Evaluation separates detection, classification, segmentation, depth error, surface error, size-estimation error, and reconstruction stability.
We also consider clinical usefulness, reduction in diagnostic time, missed-lesion reduction, and measurement reproducibility.
Current Questions
- 13D shape and absolute-size estimation of endoscopic polyps using vascular structure and optical cues
- 2Robust lesion detection, classification, and segmentation across small labels, institutions, and devices
- 3Multitask diagnostic support integrating detection, reconstruction, and malignancy estimation
- 4Learning and validation pipelines using generative models and explainability
Demo
A demo for reconstructing 3D shape from a single image is available.
Demo