Skin Cancer Kills When It Goes Undetected. Effat University Researchers Are Training Computers to Change That.

Two papers co-authored by Effat University researchers advance the use of AI in cancer diagnosis — one mapping the best available techniques, one proposing a new method for breast cancer detection.

The gap between a cancer that kills and one that doesn't is often not a question of treatment. It is a question of timing. For skin cancer, survival rates climb as high as 95% when the disease is caught and treated early. When it isn't, the outcome is far worse. That single fact — the enormous difference that early detection makes — is what gives the current wave of AI research in oncology its urgency.

The clinical bottleneck is well understood. Determining whether a tumour is benign or malignant requires a specialist. Dermatologists with the training and experience to make that determination reliably are not available everywhere, and in regions with high patient demand or limited healthcare infrastructure, the wait for that expertise can be long enough to matter. AI does not solve every problem in cancer diagnosis, but it addresses this one directly: a model trained on thousands of tumour images can analyse a new one and classify it accurately, without the geographic and logistical constraints that limit access to human specialists.

Researchers at Effat University in Jeddah are contributing to this work on two fronts — one focused on understanding which AI techniques currently perform best for skin cancer diagnosis, and one proposing a new method for applying deep learning to breast cancer detection.

Seventeen Techniques, One Clear Frontrunner

A paper co-authored by Effat University's Saeed Mian Qaisar surveys the landscape of machine learning and deep learning approaches currently available for skin cancer diagnosis, comparing 17 techniques across the range of options the field offers.

The comparison spans methods with very different histories and design philosophies. Support Vector Machines, developed in the 1990s, are included for their well-documented accuracy. K-means Clustering and K-nearest Neighbours — techniques that date to the 1960s — are assessed for their flexibility and continued relevance. The paper also covers a range of deep learning models, including Long Short-Term Memory networks and Deep Neural Networks.

The standout finding is the performance of Convolutional Neural Networks. CNNs — deep learning models designed specifically for image analysis — have demonstrated accuracy above 90% in predicting different types of skin cancer, making them the most reliable tool currently available for this application. For anyone building or deploying AI diagnostic systems for skin cancer, the paper's comparison of 17 techniques provides a clear-eyed guide to where the field currently stands and which approaches are most worth investing in.

A New Way of Looking at Breast Cancer Images

The second contribution from Effat University moves beyond reviewing existing techniques into proposing something new. A paper co-authored by researcher Abdulhamit Subasi introduces what it calls a grid-based deep feature generator — a method for analysing ultrasound images of suspected breast cancer that works differently from conventional approaches.

Rather than processing an ultrasound image as a single input, the technique divides the image into a grid of rows and columns and applies pre-trained CNN models to each section individually. The result is a more granular extraction of diagnostic features from the image — a structured way of ensuring that the model captures detail that a whole-image approach might miss.

The practical significance of this is considerable. Ultrasound imaging is already among the more accessible diagnostic modalities, making it a natural candidate for AI-assisted analysis in settings where access to specialist healthcare is constrained. A technique that improves the diagnostic information extractable from ultrasound images without requiring specialist interpretation could extend reliable breast cancer detection to populations and regions that currently lack it.

Limitations the Field Has to Confront

Both lines of research sit within a field that, despite its progress, still has significant problems to work through. The most pressing is the lack of clinical data representing all skin types in the datasets used to train AI diagnostic models. A model trained predominantly on images from one population will carry biases that affect its accuracy for patients from underrepresented groups — a limitation with direct consequences for the populations who, in many cases, have the most to gain from better access to early detection.

Clinical adoption is the other variable. The technology's value depends entirely on whether it is integrated into clinical practice, and that integration requires dermatologists and other specialists to treat AI tools as complements to their expertise rather than competitors. The evidence from the research suggests that this framing — AI as an extension of specialist capability rather than a replacement for it — is the right one, both technically and practically. Getting clinicians to embrace it is as important as getting the algorithms right.