Segmentation and classification of thyroid follicular neoplasm using cascaded convolutional neural network

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Thyroid follicular neoplasms, including follicular adenoma (TFA) and follicular thyroid carcinoma (FTC), present similar imaging characteristics in ultrasound.
• Distinguishing between TFA and FTC is challenging for specialists due to subtle differences in shape, size, and contrast.
• Segmentation of lesions from complex thyroid ultrasound images with weak edges and heterogeneous regions is a prerequisite for accurate classification.

Purpose of the Study:

• To develop and evaluate a combined segmentation and classification method for thyroid follicular neoplasms using ultrasound images.
• To accurately discriminate between thyroid follicular adenoma (TFA) and follicular thyroid carcinoma (FTC).
• To overcome challenges in lesion segmentation and classification accuracy caused by image complexity and feature extraction.

Main Methods:

• A cascaded learning architecture integrating a prior-based level set method with a deep convolutional neural network (Res-U-net) was employed for segmentation.
• The classification network shared shallow layers with the segmentation network to leverage extracted features.
• The combined approach aimed to enhance segmentation accuracy and improve classification performance.

Main Results:

• The proposed method achieved a Dice score of 92.65% for segmenting lesion areas in thyroid ultrasound images.
• The classification accuracy for distinguishing between TFA and FTC reached 96.00%.
• The cascaded architecture effectively addressed segmentation and classification challenges.

Conclusions:

• The developed method demonstrates high accuracy in segmenting thyroid lesions and classifying follicular neoplasms.
• This AI-driven approach shows significant potential for improving the diagnosis of thyroid follicular adenoma and carcinoma.
• The combination of advanced segmentation techniques and deep learning offers a promising solution for challenging medical image analysis tasks.