Pranjal Sahu, Ph.D. Thesis Defense: 'Novel Machine-learning-centric Data Synthesis Algorithms and Analysis Techniques for Medical Imaging'

Friday, May 7, 2021 - 2:00pm to 4:00pm
Zoom - contact for Zoom info.
Event Description: 




Deep learning has shown tremendous success in solving some of the long-standing computer vision problems and has found applications in the medical imaging domain as well. Unlike natural light computer vision problems, medical imaging poses additional challenges such as lack of training data, the requirement of domain experts for annotation, etc. To circumvent such problems, highly reliable and novel methodologies that can tackle problems in multiple modalities are urgently needed.  With a focus on cancer diagnosis, in this thesis, we propose novel data utilization techniques in solving problems such as biomarker detection, image segmentation, de-noising, and volume reconstruction.

First, for improving Digital Breast Tomosynthesis (DBT) image acquisition, we propose novel CNN-based models for projection de-noising and scatter correction.  Both models are trained entirely using synthetic data which obviates the need for real data. The volume reconstructed from de-noised projections has enhanced calcification visibility and shows a superior signal-to-noise ratio. And the scatter corrected projection images show reduced cupping artifacts. After image acquisition, at the volume reconstruction stage, we propose a CNN model to provide an interactive capability to the practitioner that can help in on-the-fly tuning of the image quality.  The proposed CNN model works on both DBT and CT datasets and can improve the workflow of practitioners.

Second, for segmenting the lung volume from reconstructed CT volumes we propose a novel structure correction method. The proposed method solves the problem of the inclusion of tumors/large nodules in the lung segmentation masks.  Typically training a segmentation CNN would require human-annotated segmentation masks including the tumor region, however, in our proposed method we achieve the desired results without requiring fully annotated lung segmentation masks.  We re-purpose the already available nodule annotations to generate synthetic tumor corruption masks which are used to generate corrupted segmentation masks.  A 3D CNN is then trained using the generated data which performs the structure correction.

Third, we propose a novel and lightweight multi-section CNN for lung nodule classification.  A novel data utilization technique is adopted here which aggregates the information from multiple cross-sections of a 3D lung nodule and achieves state-of-the-art results for classification tasks on the LIDC-IDRI dataset. The model is a 2D CNN but due to the multiview samplings, it is able to extract the volumetric information from a 3D lung nodule. Being lightweight multi-section CNN can be ported to a tablet and brings the benefit of AI-driven applications into practitioners’ hands.

Fourth, we propose a hand-held computer (smart-phone, Raspberry Pi) based assistant that classifies with the dermatologist-level accuracy skin lesion images into malignant and benign and works in a standalone mobile device without requiring network connectivity. The proposed model utilizes a hybrid approach based on advanced deep learning model and domain-specific knowledge and features that dermatologists use for the inspection purpose to improve the accuracy of classification between benign and malignant skin lesions. The experiments conducted on the ISIC 2017 skin cancer classification challenge demonstrate the effectiveness and complementary nature of these hybrid features over the standard deep features.

At the end, we will also highlight the work on a similar topic of semi-supervised learning for covid lesion segmentation from CT scans. Also we will discuss the future work in the medical image reconstruction domain to automatically adapt the image quality for optimum reader performance.

Computed Event Type: