From skin cancer classification to a new grid-based approach to breast cancer diagnosis, researchers at the Jeddah institution are contributing to one of the most consequential applications of machine learning in medicine.
Early detection is the single most important factor in cancer survival. For skin cancer specifically, the numbers are stark: while the disease carries an alarmingly high mortality rate when caught late, survival rates reach up to 95% when it is identified and treated early enough. The difference between those two outcomes often comes down to access — access to a specialist who can look at a tumour and determine whether it is benign or malignant, and access to that specialist quickly enough to matter.
That access is not evenly distributed. Skilled dermatologists are in short supply in many parts of the world, and the time required for thorough clinical analysis creates bottlenecks that are most severe in regions with high patient loads or limited medical infrastructure. For patients in those regions, the gap between what early detection could offer and what the healthcare system can deliver is wide and, under current conditions, difficult to close.
Artificial intelligence offers a way to address that gap directly. Machine learning models — computers trained on libraries of tumour images to recognise and classify patterns — can analyse a new image and sort it into the correct diagnostic category without requiring a specialist to be physically present or available. At Effat University, researchers are actively contributing to this field, both by mapping the current landscape of available techniques and by developing new approaches that push diagnostic accuracy further.
Mapping the State of the Art
A recent paper co-authored by Effat University’s Saeed Mian Qaisar provides a comprehensive comparison of the machine learning and deep learning techniques currently available for skin cancer diagnosis, covering 17 distinct methods and evaluating their relative strengths.
The review covers a range of approaches, from established techniques like Support Vector Machines — developed in the 1990s and noted for their high accuracy — to more flexible methods such as K-means Clustering and K-nearest Neighbours, which date to the 1960s but remain widely used. The findings point clearly toward deep learning models as the current performance leaders. Among these, Convolutional Neural Networks stand out — CNNs have demonstrated the ability to accurately predict different types of skin cancer with accuracy exceeding 90%, making them the most reliable tool the field currently has for image-based diagnostic classification.
Other deep learning architectures covered in the review include Long Short-Term Memory networks and Deep Neural Networks, each with their own characteristics and use cases within the broader diagnostic landscape.
A New Approach to Breast Cancer Diagnosis
Alongside the skin cancer review, Effat University researcher Abdulhamit Subasi has co-authored a paper proposing a novel method for AI-assisted breast cancer diagnosis using ultrasound images.
The technique — described as a grid-based deep feature generator — works by dividing an image of suspected breast cancer into rows and columns, then applying pre-trained CNN models to each segment individually. By analysing the image in structured sections rather than as a single whole, the approach is designed to extract richer and more precise diagnostic information than conventional methods allow.
The practical value of approaches like this extends well beyond the technical. Ultrasound-based diagnosis is already more accessible than some other imaging modalities, and AI tools that can extract reliable diagnostic information from those images without requiring specialist interpretation could meaningfully expand access to accurate breast cancer detection in settings where radiologists and oncologists are not readily available.
What Still Needs to Be Done
The research coming out of Effat University reflects both the genuine progress the field has made and the work that remains. One significant limitation acknowledged in the current state of AI cancer diagnosis is the lack of clinical data representing all skin types. Training models on datasets that do not reflect the full diversity of the patient population introduces biases that can affect diagnostic accuracy for underrepresented groups — a problem that has direct health equity implications and that the field needs to address as a priority.
There is also the question of clinical adoption. AI diagnostic tools are only useful if they are actually used, and that requires dermatologists and other clinicians to engage with them as complementary instruments rather than viewing them as a threat to professional practice. The technology works best when it functions alongside clinical expertise, extending the reach of specialists rather than attempting to replace them.
