AI Breakthrough in Early Melanoma Detection
The Growing Threat of Melanoma
According to the American Academy of Dermatology, an estimated 212,000 new cases of melanoma, the most severe type of skin cancer, are expected to be diagnosed in the U.S. in 2025. The danger of melanoma lies in its ability to spread to lymph nodes and internal organs if not detected early, making it a life-threatening condition. Early and accurate detection is therefore critical to saving lives.
A New AI-Powered Solution
To tackle this challenge, researchers from Northeastern University have developed a groundbreaking artificial intelligence system. Divya Chaudhary, an assistant teaching professor, and Peng Zhang, a graduate student, created a highly efficient hybrid system named the SegFusion Framework. This tool is designed to assist medical professionals in identifying melanoma with greater speed and precision.
"If we can detect it early, we can save a lot of lives and help medical practitioners, clinicians in early diagnosis," explains Chaudhary.
How the Hybrid System Works
SegFusion ingeniously combines the strengths of two powerful deep learning models. The first model, U-Net, excels at segmentation—it scans skin images and draws boundaries around any suspicious-looking regions. The second model, EfficientNet, then analyzes these highlighted areas to classify them as either cancerous or benign.
To train this sophisticated system, the researchers utilized two extensive dermatology image collections. The HAM10000 dataset, with over 10,000 images of pigmented lesions, was used to train the segmentation model. For the classification task, the team used the ISIC 2020 dataset, which contains over 33,000 images labeled as melanoma or non-melanoma. Since melanoma cases constituted only 1.8% of this dataset, the team balanced it by oversampling positive cases and undersampling negative ones, ensuring the AI learned to identify both conditions effectively.
A custom "data bridge" was built to connect the two models. The segmentation model creates a black-and-white mask of a suspicious spot, which the bridge then overlays onto the original image. This allows the classification model to focus its analysis on the precise area of concern.
Proven Performance and Unprecedented Accuracy
When benchmarked against other popular AI methods, the SegFusion Framework demonstrated superior performance. On the International Skin Imaging Collaboration 2020 dataset, it achieved an impressive 99.01% accuracy in correctly identifying melanoma. This significantly surpasses the performance of four other machine learning approaches, including ResNet-101+SVM (97.15%), NasNet (97.7%), InSiNet+U-Net (90.54%), and MobileNetV2 (98.2%).
Divya Chaudhary, an assistant teaching professor of computer science at Northeastern’s Seattle campus, says SegFusion could be adapted to detect other cancers, such as breast or lung cancer. Courtesy photo
The Future of AI in Cancer Diagnosis
Currently, the process involves some manual steps between the two models. However, the team's ultimate goal is to merge them into a single, fully automated system that streamlines the entire process from image capture to diagnosis.
Looking ahead, Chaudhary and her students plan to enhance the system's accuracy even further by integrating patients' health records, such as blood pressure and oxygen levels. They also envision creating a dedicated app for dermatologists, where the AI can work in the background to provide real-time assistance during patient examinations.
The potential of this technology extends far beyond skin cancer. Chaudhary believes the SegFusion framework could be adapted to help detect other malignancies, including breast and lung cancer.
"The students are working with me on building a big cancerous framework," she says. "You will be able to put any picture into several AI models we are building, and it will clearly tell you."
