Towards Finding the 'Needle in the Haystack': DeepHealth's Entry to the Mammography DREAM Challenge

Abstract

While computer vision models designed for natural object recognition are often successful in direct transfer learning to other domains, the high resolution and local basis of decision-making innate in mammogram classification presents a unique challenge. This was particularly relevant in The Digital Mammography DREAM Challenge, where only image-level labels were available for training and the dataset was highly imbalanced. To overcome these difficulties, we implemented a two-stage training scheme consisting of first training a patch classifier using a popular public dataset containing lesion segmentation masks, followed by image-level training on the public and DREAM datasets. The image-level model was instantiated by a global aggregation of features from the patch model, used in a scanning window (i.e. convolutional) fashion. With the patch-trained initial weights, the global model trained efficiently end-to-end, achieving accuracies up to 0.87 AUROC in the competition.