Deep learning automated dermatopathology

Inventors

Jackson, Brian H. • Stavish, Coleman C.

Assignees

Proscia Inc

Abstract

Techniques for classifying a human cutaneous tissue specimen are presented. The techniques may include obtaining a computer readable image of the human tissue sample and preprocessing the image. The techniques may include applying a trained deep learning model to the image to label each of a plurality of image pixels with at least one probability representing a particular diagnosis, such that a labeled plurality of image pixels is obtained. The techniques can also include applying a trained discriminative classifier to contiguous regions of pixels defined at least in part by the labeled plurality of image pixels to obtain a specimen level diagnosis, where the specimen level diagnosis includes at least one of: basal cell carcinoma, dermal nevus, or seborrheic keratosis. The techniques can include outputting the specimen level diagnosis.

Core Innovation

The invention describes a method of classifying a human cutaneous tissue sample by obtaining an image of the sample and applying a trained machine learning system. The trained machine learning system outputs a dermatological diagnosis as the result, and the method uses training images that include both diagnostic goal dermatological images and distractor dermatological images.

The training images are organized into at least a plurality of images of a first dermatological diagnostic goal and a plurality of images of a second dermatological diagnostic goal. For each diagnostic goal, distractor images corresponding to that diagnostic goal are also provided, and the trained machine learning system is trained by training a first deep learning neural network using the plurality of images of the first dermatological diagnostic goal together with the distractor images corresponding to the first dermatological diagnostic goal, and training a second deep learning neural network using the plurality of images of the second dermatological diagnostic goal together with the distractor images corresponding to the second dermatological diagnostic goal.

In the described workflow, whole-slide image processing is used, with pixel-level labeling that associates disease probabilities to image regions. Contiguous high-probability regions are extracted, converted into feature vectors, and used to drive discriminative classification for a specimen-level dermatological diagnosis. A confidence level and threshold-based reporting behavior are used to support outputting a dermatological diagnosis, including optional reporting and visualization behaviors that reduce human intervention.

Claims Coverage

The independent claims cover two aspects of the same inventive approach: a method of classifying a human cutaneous tissue sample using a trained machine learning system trained with diagnostic goal images and distractor images split across first and second diagnostic goals, and a non-transitory computer readable medium with instructions to perform the same classification actions. Across the independent claims, the inventive core includes the two-part diagnostic-goal training with separate deep learning neural networks and the use of corresponding distractor images for each diagnostic goal, followed by outputting a dermatological diagnosis.

Both independent claims are centered on classifying a human cutaneous tissue sample by applying a trained machine learning system that is trained using diagnostic-goal dermatological images together with distractor dermatological images for a first diagnostic goal and a second diagnostic goal. The training is performed by training separate deep learning neural networks for the first and second diagnostic goals using their corresponding distractor images, and the trained system produces a dermatological diagnosis that is output.

Stated Advantages

Documented Applications