PARVATIKAR, AKASH
(2022)
Integrating AI-based applications in anatomic pathology workflows.
Doctoral Dissertation, University of Pittsburgh.
(Unpublished)
Abstract
Traditional pathological diagnosis is considered as the gold standard by clinicians. However, this manual practice can be inefficient, error-prone, and highly subjective. To mitigate these issues, digital pathology is gaining traction which has attracted researchers to build black-box AI-based approaches intended to assist anatomic pathology workflows. The success of such approaches is dependent on large-scale generation of pathologist annotated high quality training data which is a serious bottleneck in computational pathology. Additionally, the AI systems must be interpretable and minimize the time-to-decision to achieve clinical adoption and possibly facilitate regulatory agency approvals.
We hypothesize that building computational models of already established anatomic pathology knowledge will alleviate the training data generation bottleneck and develop clinically interpretable models. In addition, implementing computational pathology workflows on the emerging customizable computing AI-based architectures will satisfy high-throughput and minimal time-to-decision requirements.
In this thesis, we tested our hypothesis on differential diagnoses of breast biopsies. We invoke analytical models to provide a quantitative assessment of the structural changes in the breast tissue along a diagnostic continuum triggered by atypia and other malignancies. We further combine the analytical models with a prototype-driven learning strategy to provide interpretability and achieve a superior classification performance in diagnosing breast biopsies over the state-of-the-art methods. To showcase the potential for seamless integration of our computational pathology framework into clinical workflows, we use a next generation high performance AI-based computing architecture to detect histological structures in breast tissue and classify them as high-risk vs low-risk. A key contribution of our framework is in building a communication platform for pathologists and computational scientists to interact and develop AI-based applications and to enhance patient care in a clinical setting.
Share
| Citation/Export: |
|
| Social Networking: |
|
Details
| Item Type: |
University of Pittsburgh ETD
|
| Status: |
Unpublished |
| Creators/Authors: |
|
| ETD Committee: |
|
| Date: |
10 June 2022 |
| Date Type: |
Publication |
| Defense Date: |
25 February 2022 |
| Approval Date: |
10 June 2022 |
| Submission Date: |
29 April 2022 |
| Access Restriction: |
2 year -- Restrict access to University of Pittsburgh for a period of 2 years. |
| Number of Pages: |
107 |
| Institution: |
University of Pittsburgh |
| Schools and Programs: |
School of Medicine > Computational and Systems Biology |
| Degree: |
PhD - Doctor of Philosophy |
| Thesis Type: |
Doctoral Dissertation |
| Refereed: |
Yes |
| Uncontrolled Keywords: |
breast lesions, prototype-based recognition, diagnostic explanation, digital and computational pathology |
| Date Deposited: |
10 Jun 2022 13:28 |
| Last Modified: |
10 Jun 2024 05:15 |
| URI: |
http://d-scholarship.pitt.edu/id/eprint/42736 |
Metrics
Monthly Views for the past 3 years
Plum Analytics
Actions (login required)
 |
View Item |
![[feed]](/images/feed-icon-32x32.png){kind=icon}