Abstract
Pituitary microadenoma (PM) is often difficult to detect by MR imaging alone. We employed a computer-aided PM diagnosis (PM-CAD) system based on deep learning to assist radiologists in clinical workflow. We enrolled 1,228 participants and stratified into 3 non-overlapping cohorts for training, validation and testing purposes. Our PM-CAD system outperformed 6 existing established convolutional neural network models for detection of PM. In test dataset, diagnostic accuracy of PM-CAD system was comparable to radiologists with > 10 years of professional expertise (94% versus 95%). The diagnostic accuracy in internal and external dataset was 94% and 90%, respectively. Importantly, PM-CAD system detected the presence of PM that had been previously misdiagnosed by radiologists. This is the first report showing that PM-CAD system is a viable tool for detecting PM. Our results suggest that PM-CAD system is applicable to radiology departments, especially in primary health care institutions.
Introduction
Pituitary microadenomas (PM) are small tumors, defined as being less than 10mm in diameter, located within the pituitary gland. While typically benign, their presence can trigger severe hormonal imbalances, leading to systemic conditions such as Cushing’s disease, acromegaly, or infertility.
Detecting these lesions presents a significant clinical challenge. Due to their minute size and the anatomical complexity of the sellar region, PMs are often described as a "needle in a haystack" on standard MRI scans. Diagnosis is highly subjective, relying heavily on a radiologist’s specific experience and the time available to scrutinize subtle textural changes. In high-pressure clinical environments or primary care settings, these fine-grained features are easily overlooked, resulting in delayed intervention. To address this, researchers developed PM-CAD, a deep learning-based computer-aided diagnosis system designed to provide expert-level detection.
Methodology: The PM-CAD Framework
The PM-CAD system utilizes a two-stage pipeline to mimic and then enhance the human diagnostic process:
1. Pituitary Localization
The first stage employs a Faster R-CNN framework with a ResNet-50 Feature Pyramid Network (FPN). This model is tasked with identifying and "cropping" the pituitary gland from the full brain MRI. By narrowing the search space to the region of interest (ROI), the system reduces computational noise and allows the subsequent diagnostic model to focus exclusively on relevant tissue.
2. Fine-Grained Diagnosis (PM-Net)
The core of the system is PM-Net, a modified ResNet-18 architecture. Standard convolutional neural networks (CNNs) typically use aggressive downsampling—large strides and max-pooling layers—to reduce memory usage. However, in medical imaging, this process often destroys the subtle, high-resolution textures necessary to identify a microadenoma. The researchers modified the architecture by:
- Replacing large initial kernels with 3x3 convolutions.
- Removing several pooling layers to preserve spatial resolution.
- Integrating a spatial attention module that allows the network to prioritize discriminative regions within the pituitary gland over normal tissue.
3. Robustness and Pre-processing
To ensure the model could function across different hospital settings, the team utilized histogram matching normalization and intensity shift data augmentation. These techniques help the model remain robust against variations in MRI hardware (e.g., GE vs. Philips) and scanning protocols.
Experimental Design and Datasets
The study involved 1,228 participants and a total of 28,839 images collected between 2014 and 2019. The data was divided into three distinct cohorts:
- Internal Dataset (Test A): Sourced from The Third Affiliated Hospital of Sun Yat-Sen University for training and primary testing.
- External Dataset (Test C): Sourced from Sun Yat-Sen Memorial Hospital and Harbin Medical University to evaluate the model's ability to generalize to unseen environments.
- The Misdiagnosis Cohort (Test B): A critical set of three patients who were clinically confirmed to have PMs, but whose tumors had been missed by radiologists during initial screenings.
Key Results and Discussion
Performance Against Standard Models
PM-CAD demonstrated superior performance compared to six established CNN architectures, including GoogLeNet, DenseNet, and 3D-CNN. The system achieved an Area Under the Curve (AUC) of 0.9813, indicating high sensitivity and specificity.
Human-AI Comparison
In a head-to-head evaluation against human experts, PM-CAD’s accuracy reached 94%. This performance was comparable to senior radiologists with over a decade of experience (95%) and significantly outperformed junior radiologists. Furthermore, the system’s weighted error rate was only 10%, whereas the average error rate among six human radiologists was 21.67%.
The "Safety Net" Effect
Perhaps the most significant finding came from the misdiagnosis cohort (Test B). PM-CAD successfully identified the tumors in all three cases that had been previously missed by human clinicians. This suggests that the AI can serve as a vital secondary check, catching subtle lesions that might be overlooked due to human fatigue or lack of specialized expertise.
Generalization and Limitations
While the system maintained high accuracy (90%) on external datasets, the slight drop from internal performance (94%) highlights the ongoing challenge of "domain shift" in medical AI. Different hardware and imaging protocols remain a hurdle for universal deployment. Additionally, as a retrospective study, the real-world utility of PM-CAD must still be validated through prospective clinical trials where the AI is integrated into live workflows.
Conclusion
The PM-CAD system represents a significant step forward in the automated detection of pituitary microadenomas. By prioritizing the preservation of high-resolution image features and providing a stable, fatigue-free diagnostic tool, the system offers a viable means to improve diagnostic accuracy, particularly in primary healthcare institutions where specialized neuroradiology expertise may be limited.
Summary updated