Classification of melanocytic lesions in selected and whole-slide images via convolutional neural networks

This indicates challenges associated with large and complex datasets, suggesting AI models should perform better with uncurated datasets.

“This indicates the challenges associated with large and complex datasets, suggesting that AI models should be able to perform better with uncurated datasets.”

Hart et al. evaluated a CNN to categorize melanocytic lesions; CNNs trained on carefully selected samples achieved 99% accuracy, but accuracy reduced to 92% on whole-slide images, with sensitivity 85% and specificity 99%.

“Hart et al. (22) evaluated CNN to categorize melanocytic lesions. The present study noted that when the CNNs were trained on carefully selected samples, they were able to achieve 99% in terms of accuracy rates; however, the accuracy rate reduced to 92% when the test was conducted on WSIs, with a sensitivity of 85% and a specificity of 99%.”

Collectively, the Laohawetwanit and Hart studies demonstrate a continuous pattern: models exhibit robust performance under optimum or curated settings but decline under real-world variability or whole-slide complexity.

“Collectively, these two studies demonstrate a continuous pattern in the literature: models exhibit robust performance under optimum or curated settings but experience significant declines when subjected to real-world variability or whole-slide complexity.”

The performance decline identified by Olsen et al. confirms the generalizability gap noted in Hart et al., suggesting this constraint is systemic among CNN-based models rather than an isolated occurrence.

“The performance decline identified by, Olsen et al. (28) confirms the generalizability gap noted in Hart et al. (22) suggesting that this constraint is systemic among CNN-based models rather than an isolated occurrence.”

Hart et al. demonstrated CNNs achieved accuracy 99% on selected datasets to differentiate melanocytic lesions, but this fell to 92% when using WSIs, with sensitivity 85% and specificity 99%.

“Similarly, Hart et al. (22) demonstrated that CNNs achieved an accuracy of 99% on the selected datasets to differentiate between the melanocytic lesions, but this fell to 92% when using WSIs, with a sensitivity of 85% and specificity of 99%.”

The scoping review captures how model performance reduces when datasets used are non-curated or in real-world scenarios, as seen in Hart et al. and Olsen et al.

“The scoping review also captures how model performance reduces when the datasets used are non-curated or in real-world scenarios, as seen in the works of Hart et al. (22) and Olsen et al. (28).”