- Developing computer based algorithms for prediction of chemo-radiation therapeutic response from texture based radiomic features extracted from pre-treatment MRI scans of Glioblastoma.
- Investigating morpholgical associations of predictive radiomic texture features with histopathological phenotypes of H&E stained digitized tissue images to assess the tumor phenotypical traits responsible for treatment resistance.
Fig: A represents the radiomic heatmaps of a responder and non-responder patients corresponding to highly predictive textural descriptor. In B, histopathological attributes associated with predictive radiomic texture descriptors are shown and C represents the box plots of top 3 radiomic features to predict Glioblastoma patients response to chemo-radiation therapy.
- A Dataset and a Technique for Generalized Nuclear Segmentation for Computational Pathology
Nuclear segmentation in digital microscopic images can enable extraction of high quality features for nuclear morphometric and other analysis in computational pathology. However, conventional image processing techniques such as Otsu and watershed segmentation do not work effectively on challenging cases such as chromatin-sparse and crowded nuclei. We developed a deep learning based generalized nuclear segmentation algorithm and also released a benchmark nuclear segmentation datasetencompassing multiple organs, patients and disease states.
Fig: Examples of sub-images taken from the test images for different organs (columns) showing challenging cases based on variation in nuclear appearance and crowding.
- Detecting multiple sub-types of breast cancer in a single patient
I developed a method to detect the presence of multiple aggressive molecular sub-types HER2 and basal-like in a single breast cancer patient by using computer vision on just H&E stained images. This has treatment planning implications, because the frontline therapy for breast cancer patients with HER2 gene amplification is trastuzumab (Herceptin), but it fails in roughly half of the patients. Presence of another molecular subtype, particularly basal-like, has been implicated in the failure of trastuzumab. Predicting whether a therapy will fail is of tremendous clinical importance for treatment planning.
Fig: Scatter plots of top two differentiating molecular features for HER2 vs. basal-like IDC using four technologies. Patients that seem misclassified in at least one plot (called “borderline” HER2 and basal-like in the text) have black bounding boxes.