Introduction & Objectives

The early detection of skin cancer, including melanoma, substantially improves the five-year survival rate of patients. Deep Learning has shown its capability to assess pigmented skin lesions with a level of accuracy that matches that of dermatologists. These artificial neural networks analyze images at the pixel level as they pass through various layers of the network with distinct graphic filters. However, it is important to acknowledge the limitations of such neural networks. They may struggle with rare entities due to a lack of training images or image artifacts. The best results may be achieved through a collaborative approach, combining dermatologists’ expertise with machine capabilities.

Materials & Methods

In our study, 207 anonymous cases from outpatient dermatology clinics, with lesions photographed by dermoscopy, were assessed. 370 other dermoscopic skin lesions were collected for a total of 577 images. Our algorithm achieved excellent performance in melanoma classification using dermoscopy images and few high resolution zoomed clinical images, showing a robust performance. The achievement of this score was made possible through the application of an optimization strategy, which enabled us to create a “doubtful” category derived from the two original categories: melanoma and non-melanoma. Although maintaining all three categories yielded an acceptable score, the exclusion of probabilities within the doubtful category significantly improved performance. This method is likely to be precise because it highlights, via the “doubtful” category, which lesions require the insight of human experts and which ones are clear-cut cases. This enables specialists to concentrate solely on instances that demand their full attention.

Results

The algorithm has demonstrated excellent performance, with an AUC of 95%, sensitivity of 98%, and specificity of 88% on test data.

These results also demonstrate the potential to aid primary care providers in making more informed referrals by distinguishing between pigmented lesions with a high likelihood of being melanoma and those that are probably benign, using dermoscopic images of the skin. Lesions considered ‘doubtful’ should be managed using the usual care pathway.

Conclusion

The algorithm designed for melanoma skin cancer classification has demonstrated robust performance, particularly in differentiating between extreme classes. The ability of the HUVY algorithm to identify its own limit (doubtful area) is an opportunity to mitigate False Positives and False Negatives and to alert the healthcare professional when a lesion needs human expertise. It illustrates a good complementarity of human and machine. The machine allows for an optimal care pathway for obvious cases (whether it is malign or benign) but also alerting the human when it is not obvious and that other parameters have to be taken into account. To build upon these promising results, a heavier neural network architecture could be employed to increase the model’s capacity for feature extraction and representation. Such architectures would potentially enable the algorithm to discern more intricate patterns within the data, which is especially beneficial for complex medical imaging tasks. However, this approach should be balanced against the increased computational demands and the potential for overfitting, necessitating a larger dataset for training and validation.

EADV Congress 2024, Amsterdam.