Deep learning Lyme disease diagnosis

Inventors

Billings, Seth D. • Burlina, Philippe M. • JOSHI, Neil J. • Aucott, John N. • NG, Elise

Abstract

Techniques for diagnosing Lyme disease are presented. The techniques may include obtaining a digital photo of a skin lesion, providing the digital photo to a deep learning convolutional neural network, such that an output diagnosis is provided. The deep learning convolutional neural network may be trained using a training corpus including a plurality of digital training images annotated according to one of a plurality of training image diagnoses, where the plurality of training image diagnoses include at least one Lyme disease type, normal skin, and at least one non-Lyme skin lesion type, and where the plurality of digital training images include multiple digital photographs publicly available on the internet. The techniques can include outputting the output diagnosis.

Core Innovation

The system diagnoses Lyme disease by obtaining, using a smart phone camera, a digital photo of a skin lesion and providing the digital photo to a deep learning convolutional neural network that provides an output diagnosis. The output diagnosis is one of “Lymes” or “not Lymes” and is outputting in electronic format. The approach is based on a deep learning model trained using a training corpus of digital training images annotated according to training image diagnoses.

The training corpus includes at least one Lyme disease type diagnosis, normal skin, and at least one non-Lyme skin lesion type diagnosis. The Lyme disease skin lesion training images depict simple erythema migrans skin lesions and diffuse erythema migrans skin lesions. The non-Lyme skin lesion type includes herpes zoster and further includes Tinea corporis.

The plurality of digital training images comprise multiple digital photographs publicly available on the internet. The diagnostic system supports prescreening or diagnosis by classifying lesions for Lyme disease relevance and producing an output diagnosis for electronic delivery. The model is described as a DCNN using ResNet50, applied to images captured by a smartphone or camera for on-device or remote inference.

Claims Coverage

The document presents one independent claim and dependent claims refining specific execution and training-data sourcing features. The independent claim centers on smartphone image acquisition, deep learning diagnosis into a two-label output, and a training corpus using annotated Lyme types, normal skin, and specified non-Lyme lesions obtained from publicly available internet photographs.

Overall, the claim coverage is directed to a system that performs Lyme disease diagnosis from a smartphone-captured skin lesion image using a deep learning convolutional neural network trained on annotated internet-mined photographs, with explicit inclusion of Lyme lesion types, normal skin, and non-Lyme lesion types such as herpes zoster and Tinea corporis.

Stated Advantages

Documented Applications