dSPIC: a deep SPECT image classification network for automated multi-disease, multi-lesion diagnosis

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Single-photon emission computed tomography (SPECT) bone scintigraphy is crucial for diagnosing and monitoring diseases like metastasis.
• However, SPECT imaging suffers from low resolution, poor signal-to-noise ratio, and low specificity, hindering accurate diagnosis.
• Visually similar lesions across different diseases complicate interpretation of SPECT images.

Purpose of the Study:

• To develop an automated diagnostic tool for whole-body SPECT scintigraphic images using a novel convolutional neural network.
• To enhance the classification accuracy of diseases, including metastasis and arthritis, from SPECT images.
• To address the challenge of limited SPECT image samples through advanced data augmentation techniques.

Main Methods:

• A self-defined end-to-end deep learning network, dSPIC, was developed for SPECT image classification.
• Data preprocessing involved data augmentation using geometric transformations and generative adversarial networks to expand the dataset.
• The dSPIC network was trained to extract optimal features for classifying images into normal, metastasis, arthritis, or multiple disease categories.

Main Results:

• The dSPIC network achieved high performance metrics on real-world whole-body SPECT images.
• Best performance values included accuracy (0.7747), precision (0.7883), sensitivity (0.7863), specificity (0.8820), and F-1 score (0.7860).
• Performance was evaluated on both original and augmented datasets, demonstrating robustness.

Conclusions:

• The developed dSPIC deep learning network shows significant promise for multi-disease and multi-lesion classification in SPECT bone scintigraphy.
• Its performance surpasses that of classical deep learning models like AlexNet.
• dSPIC represents a workable and effective solution for improving automated diagnosis in nuclear medicine imaging.