An ultimate goal of our research is to develop computational methods for improving image-based detection and diagnosis of disease. We are developing machine learning tools for detection, segmentation and classification of lesions in neuro- and whole-body multi-modal imaging.
Universal Lesion Detection
This project aims to develop a universal approach for lesion detection in medical scans. In a recent work presented at MICCAI 2019, we improve a RetinaNet architecture for the challenging task of small lesion detection in CT across the whole body. Our one-stage lesion detector achieves a sensitivity of 90.77% at 4 false positives per image, significantly outperforming the best reported methods by over 5%. Adding dense supervision from weak RECIST labels (lesion diameters) plus attention, improves sensitivity for small lesions <10 mm by over 8%. We obtain an almost 10% overall improvement at 0.5 FP, evaluated on DeepLesion benchmark containing a total of 32,735 lesions across the body. Can you beat our AI? Try Spot-the-Lesion!
Segmentation of Brain Lesions
Traumatic brain injury (TBI) results from brain damage caused by sudden trauma. TBI is a severe health problem and common cause of permanent disability. A promising direction for assessing TBI quantitatively is through multi-modal volumetric imaging. In particular, structural change within a short period of time is believed to be an important clinical indicator. In order to measure such change, it is essential to be able to detect and segment the pathological parts of the brain. We are working on automatic image analysis methods which allow objective and quantitative assessment of TBI lesions. This work is partially funded under the 7th Framework Programme by the European Commission (CENTER-TBI).
The methodology is applicable to other brain lesion analysis tasks, as demonstrated in our successes of winning of the Ischemic Stroke Lesion Segmentation Challenge 2015 and the Multimodal Brain Tumor Segmentation Challenges 2017.