JMIR Publications

ABSTRACT

Background:

Breast cancer-related lymphedema (BCRL) develops in 16.6% of patients after breast cancer treatment. Translational cancer laboratories study BCRL in murine tail and hindlimb lymphedema models. Manual counting and measurement of the histologic images these models produce is tedious, observer dependent, and unreliable when image quality varies. Artificial intelligence (AI) may quantify these images more objectively, but laboratories differ in the technical resources available to adopt it.

Objective:

This study aimed to describe 2 complementary categories of AI integration for translational cancer laboratories and to provide an example from our BCRL laboratory.

Methods:

The first category was annotation-driven, no-code software (HALO AI and Imaris), in which investigator annotations train the software to batch process images without programming. The second was custom-coded machine learning and deep learning, in which models learn image features from labeled images of normal and lymphedematous tissue. As a practical example, epidermal nuclei were counted in 3 hematoxylin and eosin-stained skin specimens from a murine tail lymphedema model by a custom-coded classifier (software trained on investigator-marked nuclei), by HALO AI, and manually. Manual counts were estimated by counting 1 region of each image and multiplying by the number of equivalent regions.

Results:

Nuclear counts differed between methods: the custom classifier counted 1,477, 1,544, and 3,830 nuclei in the 3 specimens, HALO AI counted 2,406, 3,700, and 3,586, and manual counts were 2,700, 4,000, and 2,650. HALO AI counts ranged from 89.1% to 135.3% of manual counts. Custom classifier counts ranged from 38.6% to 144.5%.

Conclusions:

Annotation-driven no-code platforms and custom-coded machine learning workflows are complementary approaches that translational cancer laboratories can adopt according to their technical resources. AI-assisted counts require validation against investigator measurements before use in definitive quantitative conclusions. AI-assisted image analysis may improve the objectivity, reproducibility, and scalability of translational BCRL research.