Fig 1.
A schematic diagram of collected data.
Table 1.
Detailed demographic information of the collected datasets.
Fig 2.
Training strategy for developing sensitive classification networks using GAN inversion as a feature-extracting method.
The first step is an upstream task that trains the generative adversarial network (GAN). The second step classifies through linear probing and projection combination using GAN inversion.
Fig 3.
False-negative case examples in the (A) internal test and (B) external validation datasets.
Table 2.
Our final model’s scoliosis classification performance.
Fig 4.
Scoliosis classification performance within the number of projection head layers and downstream training samples.
Fig 5.
Downstream validations relative to upstream pre-trained weight type.
Table 3.
Performance metric comparison under a fixed 0.9 sensitivity and statistical analysis using independent t-test among different feature extractor weights.
Fig 6.
Example of a well-reconstructed normal sample.
(Left) original chest X-ray. (Right) Reconstructed image using the GAN inversion technique. The encoding generator was only trained with chest X-ray images with scoliosis.
Table 4.
Image-reconstruction quality metrics of scoliosis and normal downstream training sets.
Fig 7.
(Left) Samples from a model trained on an imbalanced dataset (our method). (Right) Samples from a model trained on a balanced dataset. Red box: a sample radiograph that can be diagnosed as a normal spine in generated samples.
Fig 8.
Two manipulated image examples.
Note that scoliosis forms traverse between levoscoliosis and dextroscoliosis.