A Comparative Assessment of Different Approaches of Segmentation and Classification Methods on Childhood Medulloblastoma Images

Computational pathology involves the analysis of pathological images at two powers of microscopic examination: low (or architectural) power and high (or cell) power. Analysis at both these levels is highly crucial for treatment planning, or prognosis, of the patient. The present paper is a study on childhood medulloblastoma (CMB) using an indigenously collected image dataset. The region of interest (RoI) for the low power is patches (or sections) from the architectural level and for the high power, the nucleus. Four deep learning semantic segmentation and eight machine learning segmentation algorithms were compared and evaluated on the same dataset. The performance was measured using the Jaccard coefficient, which established the superiority of Fractal Net with 79.21% over other algorithms. Metrics such as Accuracy, Dice coefficient, F1-Score, Loss, Precision and Recall were used to compare the deep learning segmentation methods. Jaccard loss was used as an evaluation matrix for the traditional segmentation experiments. Subsequently, classification experiments were performed for comparison at both the powers and binary (normal vs abnormal) as well as multilevel (four subtypes of CMB) classification. The cell-based classification study showed 95.4% and 62.1% accuracy for binary and multi-level, respectively. Here, the features texture, shape, and color contributed to optimum classification. Next, the patch-based classification experiments involved a comparison of texture analysis using machine learning methods with two pre-trained deep learning classification models: Alexnet and VGG-16, using a softmax classifier. Here, it was observed that machine learning models outperform the deep learning models with 100% and 91.3% accuracy for both binary and multi-level, respectively. We hypothesize that combining both architectural and cell classification could lead to a more effective prognosis. The strength of the paper is the combined segmentation and classification study at two powers of microscope magnification using both classical machine learning as well as current deep learning techniques.