Melanoma prognosis prediction using image processing and machine learning

Abstract

Death of melanoma cancer is most common in Europe, and northern Europe has the second highest mortality rate of melanoma in the world, with 1.9 per 100 000 dying from melanoma in northern Europe in 2020.

Prognosis of melanoma is nowadays based on educated guesses by a pathologist, from analyzing patient tumors. Analyzing tumors takes much time, which limits the pathologist’s capability for the number of tumors they are able to analyze in a given amount of time.

The primary objective of this thesis is to suggest a machine learning based method for predicting prognosis of melanoma, to aid the pathologist. The proposed method is based on the VGG16 architecture with pre-trained weights as the backbone, adding some fully connected layers. The network is trained and validated on whole slide images (WSI) from 51 patients with known melanoma prognosis, produced at Stavanger University Hospital. Regions of interest (ROI) areas in these images are marked by a pathologist.

A foreground segmentation algorithm for skin histological WSI is presented. Tiles are extracted from ROI areas in the WSIs, and resized to contain three different magnification levels, which are used in model training and validation. Multiple magnification levels are used to mimic the way a pathologist analyzes tissues at different magnifications.

Experiments are done by combining different magnification scales, utilizing tiles from respectively one, two and three magnification level(s) to train the models.

The best performing model used only one magnification scale at 20x. Cross validation results gave a F1 score of 0.7667, and an area under the curve in a receiver operating characteristic curve of 0.81. This result is promising, considering the small number of patients in the dataset. For future work, the method has to be tested on a larger dataset. It is also recommended to test a larger set of possible hyperparameters and/or model architectures.

Death of melanoma cancer is most common in Europe, and northern Europe has the second highest mortality rate of melanoma in the world, with 1.9 per 100 000 dying from melanoma in northern Europe in 2020.

Prognosis of melanoma is nowadays based on educated guesses by a pathologist, from analyzing patient tumors. Analyzing tumors takes much time, which limits the pathologist’s capability for the number of tumors they are able to analyze in a given amount of time.

The primary objective of this thesis is to suggest a machine learning based method for predicting prognosis of melanoma, to aid the pathologist. The proposed method is based on the VGG16 architecture with pre-trained weights as the backbone, adding some fully connected layers. The network is trained and validated on whole slide images (WSI) from 51 patients with known melanoma prognosis, produced at Stavanger University Hospital. Regions of interest (ROI) areas in these images are marked by a pathologist.

A foreground segmentation algorithm for skin histological WSI is presented. Tiles are extracted from ROI areas in the WSIs, and resized to contain three different magnification levels, which are used in model training and validation. Multiple magnification levels are used to mimic the way a pathologist analyzes tissues at different magnifications.

Experiments are done by combining different magnification scales, utilizing tiles from respectively one, two and three magnification level(s) to train the models.

The best performing model used only one magnification scale at 20x. Cross validation results gave a F1 score of 0.7667, and an area under the curve in a receiver operating characteristic curve of 0.81. This result is promising, considering the small number of patients in the dataset. For future work, the method has to be tested on a larger dataset. It is also recommended to test a larger set of possible hyperparameters and/or model architectures.