Thyroid Carcinoma Detection on Whole Histologic Slides Using Hyperspectral Imaging and Deep Learning.

Area of Science:

• Oncology
• Medical Imaging
• Computational Pathology

Background:

• Hyperspectral imaging (HSI) is a non-invasive technique with growing applications in biological and medical fields.
• Its utility in oncology, particularly for analyzing histologic samples, is an active area of research.
• Follicular thyroid carcinoma presents diagnostic challenges in histologic evaluation.

Purpose of the Study:

• To compare the performance of deep learning image classifiers trained on different imaging modalities.
• To evaluate the efficacy of hyperspectral imaging data versus standard RGB and synthesized RGB data for cancer classification.
• To determine if HSI enhances the accuracy of automated thyroid cancer detection on whole histologic slides.

Main Methods:

• Three datasets were created from 33 fixed head and neck follicular thyroid carcinoma tissue samples: RGB, hyperspectral (HS), and HS-synthesized RGB.
• Three distinct deep learning classifiers were trained separately on each dataset.
• Performance was evaluated using the area under the receiver operator characteristic curve (AUC-ROC).

Main Results:

• The deep learning classifier trained on hyperspectral imaging data achieved an AUC-ROC of 0.966.
• This performance was superior to classifiers trained on RGB (AUC-ROC not specified) and HS-synthesized RGB data (AUC-ROC not specified).
• HSI data demonstrated a clear advantage in improving cancer classification accuracy.

Conclusions:

• Hyperspectral imaging significantly enhances the performance of deep learning models for classifying thyroid cancer on whole histologic slides.
• HSI, combined with deep learning, offers a powerful automated tool for thyroid cancer detection.
• This approach holds potential for improving diagnostic efficiency and accuracy in histopathology.