Page 101 - AI for Good Innovate for Impact

AI for Good Innovate for Impact

2 Use Case Description

2�1 Description 4.1-Healthcare

Context and Background

Brain cancer is a grave malignancy type that is a consequence of unconstrained and unrestricted
multiplication of cells in the brain, generally engendered by a malignant brain tumour.

Meningiomas are among the most common tumours of the Central Nervous System (CNS).
These sorts of tumours, which have a varied growth potential, emerge from the meninges, the
membranes that coat the brain and spinal cord, and generally do not metastasize.
Necrosis - It is a dead region, and there is an absence of nuclei. It consists of small, dust-like
structures. The cells are broken down along with cell borders, and there is a larger quantity
of cell debris. There are many cells and connective tissues are fewer. Texture changes occur
slowly in necrosis.

A biopsy involves examining a tissue under a microscope to determine the presence or extent
of disease spread.

Histopathological (biopsy) analysis is a painstaking, prolonged, and labour-intensive skilled job,
dependent on the expertise of the pathologist. With the availability of high-resolution digital
images of slides, referred to as Whole Slide Images (WSI), digital histopathology opens the
potential to automate and improve parts of this analysis process, with the goal of more efficient
and less error-prone diagnosis, and reducing the workload of pathologists.

Analysis of histopathological (biopsy) slides by pathologists is considered the gold standard
in determining the presence and nature of diseases such as cancer. Thin sections of tissue
extracted during biopsies are fixed on a glass slide and stained with specific chemicals, e.g.,
Haematoxylin and Eosin (H&E) stains, and then observed under a microscope at different levels
of magnification.

Different magnification levels are used for different types of analysis, from which are derived
the final impressions and diagnosis, such as categorization of tumours as malignant or benign,
and if cancerous, the grade of the disease. In our study, the selected glass slides, stained with
H&E, were scanned using a Leica SCN400 microscopic whole-slide scanner with nominally
20x magnification.

Aims and Objectives

It was observed that there is a lack of publicly accessible labelled datasets of meningioma whole
slide images. Hence, the first objective in the current work was the generation of a sizable,
high-quality, expert validated, ground truth dataset of pathology features of meningioma whole
slide images, which could be used for this research, and would also form the basis of future
work on neuropathology analysis.

The second and primary objective was the development of ML pipelines for detecting regions of
necrosis in whole slide images of meningioma. This includes a comparison of the performance
of various ML and ensemble algorithms, as well as the evaluation of performance metrics such
as specificity and sensitivity of such algorithms with varying sizes of necrotic regions.

65