RESUMO
BACKGROUND: Malignant Melanoma (MM) is characterized by a growing incidence and a high malignant potential. Besides well-defined prognostic factors such as tumour thickness and ulceration, the Mitotic Rate (MR) was included in the AJCC recommendations for diagnosis and treatment of MM. In daily routine, the identification of a single mitosis can be difficult on haematoxylin and eosin slides alone. Several studies showed a big inter- and intra-individual variability in detecting the MR in MM even by very experienced investigators, thus raising the question for a computer-assisted method. OBJECTIVE: The objective was to develop a software system for mitosis detection in MM on H&E slides based on machine learning for diagnostic support. METHODS: We developed a computer-aided staging support system based on image analysis and machine learning on the basis of 59 MM specimens. Our approach automatically detects tumour regions, identifies mitotic nuclei and classifies them with respect to their diagnostic relevance. A convenient user interface enables the investigator to browse through the proposed mitoses for fast and efficient diagnosing. RESULTS: A quantitative evaluation on manually labelled ground truth data revealed that the tumour region detection yields a medium spatial overlap index (dice coefficient) of 0.72. For the mitosis detection, we obtained high accuracies of above 83%. CONCLUSION: On the technical side, the developed iDermatoPath software tool provides a novel approach for mitosis detection in MM, which can be further improved using more training data such as dermatopathologist annotations. On the practical side, a first evaluation of the clinical utility was positive, albeit this approach provides most benefit for difficult cases in a research setting. Assuming all slides to be digitally processed and reported in the near future, this method could become a helpful additional tool for the pathologist.
