New algorithm represents a promising tool for the early detection and classification of cancer, with the potential to streamline the diagnostic process
For the approximately 20 million new cases of cancer reported each year, early and accurate diagnosis is crucial.聽 The demand for precise diagnostic tools has never been higher, and traditionally, the visual examination of tissue slides by pathologists has been the gold standard in cancer detection. However, with advancements in digital pathology, these tissue slides can now be digitized into multi-gigapixel whole slide images (WSIs). These high-resolution images hold immense potential for machine learning applications, but their sheer size presents a significant challenge.
A team of researchers including Khalifa University鈥檚 Dr. Sajid Javed and Prof. Naoufel Werghi, proposed an innovative, fully unsupervised machine learning approach to classify WSIs, enabling faster and more accurate cancer detection, among other clinical uses. The team also included researchers from Information Technology University, Pakistan, and the University of Warwick, United Kingdom. Their results were published in, a top 1% journal.
Classifying WSIs involves analyzing the image to determine whether it contains cancerous tissue. Deep learning models for classifying WSIs already exist but they often require manual annotations for expert pathologists, which is both time-consuming and costly. Recent advancements have introduced weakly supervised learning methods to alleviate this burden, but these still depend on large, labelled datasets. The research team鈥檚 fully unsupervised approach bypasses the need for any labelled data.
Their algorithm divides WSIs into smaller, more manageable patches. These patches are then transformed and subsequently inverse-transformed back to their original spaces. The transformation error 鈥� the difference between the original and inverse-transformed patches 鈥� is used to generate 鈥減seudo labels鈥�. This method hinges on a crucial observation: Normal tissue patches tend to be more homogenous than cancerous patches, which exhibit greater variability in texture and patterns.
The algorithm then further refines the labels to reduce noise and enhance accuracy. This mutual learning process continues iteratively, with each cycle improving the model鈥檚 performance.
The researchers tested their algorithm on four publicly available datasets, with their model outperforming existing state-of-the-art approaches in fully unsupervised settings, underscoring its potential for effective cancer diagnosis without the need for expensive and labor-intensive annotations from human experts. These implications are profound, with this algorithm a promising tool for the early detection and classification of cancer, making it more efficient, accurate and accessible.
The research team says further research could explore the integration of this method with semi-supervised or weakly supervised approaches to further enhance its accuracy and applicability in clinical settings.聽
Jade Sterling
Science Writer
11 July 2024
