Introduction
Avascular Necrosis (AVN) of the femoral head is a debilitating condition characterized by bone tissue death due to a disrupted blood supply, often leading to joint collapse. While Magnetic Resonance Imaging (MRI) is highly effective at detecting AVN in its early stages—typically appearing as low-signal regions in the bone marrow—the diagnostic process has traditionally relied entirely on human radiological expertise.
In an early-90s study, researchers at the Mayo Clinic demonstrated that a relatively simple multi-layer perceptron (MLP) could automate this process, diagnosing AVN directly from 2D MRI scans with up to 97% accuracy. This work was pioneering, pushing neural network applications in medicine beyond one-dimensional waveforms (like ECGs or EEGs) and into the realm of diagnostic radiology. The authors argued that radiology is a "natural" fit for neural networks, as it involves mapping a complex visual input to a simple diagnostic output, mirroring the supervised learning process of radiology residents.
Dataset and Preprocessing
The study utilized a high-quality clinical dataset of 125 sagittal T1-weighted MRI images of femoral heads from 40 adult patients. Approximately 51% of these images showed evidence of AVN across various stages of severity. Because the disease's appearance varies significantly with slice depth, the dataset included multiple cross-sectional slices of the same femoral head.
To make the problem tractable for the hardware of the era, the researchers applied strict preprocessing. Original 128x128 MRI crops of the femoral head were downsampled to 32x32 pixels, and the grayscale resolution was reduced from 256 to just 16 levels. Finally, input values were scaled to a range of ±0.5 before being fed into the network.
Network Architecture and Methods
The authors employed a standard feed-forward, fully connected multi-layer perceptron (MLP). The architecture consisted of 1024 input nodes (corresponding to the 32x32 flattened image), a single hidden layer ranging from 4 to 30 nodes, and 2 output nodes for binary classification (AVN positive or negative).
Instead of standard backpropagation, the network was trained using conjugate gradient optimization. This approach yielded remarkably fast convergence, typically completing in fewer than 50 epochs and requiring less than a minute of compute time on a SPARCstation 2.
Results and Insights
The model achieved impressive quantitative benchmarks, reaching 100% training accuracy across various configurations.
Key insights from the experiments included:
- Linear Separability: Surprisingly, the dataset was largely linearly separable. A network with zero hidden nodes (effectively a linear classifier) achieved over 91% accuracy on the test data.
- The Complexity "Sweet Spot": While linear models performed well, adding a small hidden layer optimized performance. Accuracy peaked at 97.0% using a 7-node hidden layer trained on an 80% data split. However, increasing the hidden nodes up to 30, or adding a second hidden layer, degraded performance due to overfitting on the small dataset.
- Raw Pixels vs. Engineered Features: The authors tested sophisticated engineered features, including 2-D Fast Fourier Transforms (FFTs), power spectra, and features from ring-wedge detectors. Counterintuitively, none of these outperformed raw pixel data.
- Human Baseline: When compared to humans with no formal radiology training (using the same 50% training split), the AI performed almost identically (~92% accuracy). This suggests the network learned to identify the same visual cues as a non-expert human.
Limitations and Future Directions
The authors transparently acknowledged the performance gap between their model and professional radiologists, who typically achieve near 100% accuracy by synthesizing multiple high-resolution slices and clinical context.
To bridge this gap, they proposed several future directions:
- Multi-slice Input: Developing networks capable of processing multiple images from a single patient simultaneously to emulate human volumetric analysis.
- Uncertainty Outputs: Incorporating a "not sure" classification. Radiologists often set aside ambiguous images rather than guessing, a capability the authors wished to replicate in their models.
- Detecting the Invisible: The researchers hypothesized that neural networks might eventually identify subtle, sub-visual patterns in MRIs that correlate with confirmed diagnoses but remain imperceptible to the human eye.
Summary updated