Computed tomography (CT)
What did the neural network learn?
The frozen public-data study tests whether a residual U-Net improves on transparent interpolation. Its learning curve, comparisons, compute and unsuccessful branches remain inspectable.
Four neighbouring slices → one reconstructed centre
This conceptual diagram shows the actual model family, not a count of neurons or parameters. The encoder extracts context; skip connections carry detail to the decoder; a residual correction refines the centre estimate.
The interactive CT page still runs interpolation. Opening this report does not train or load a neural checkpoint.
How the network learns and how we measured improvement
The task hides a known centre slice and provides four neighbouring slices as input. During training, the model compares its estimate with the retained centre image. It learns a correction to the linear-interpolation estimate. Two-dimensional convolutions use the four slices as channels: this is 2.5D context, rather than a full 3D acquisition generator.
The encoder learns local features, a smaller spatial representation captures context, and the decoder combines that context with fine details from a skip connection. A final addition combines the predicted residual map with the linear baseline. The interactive workbench continues to use the transparent interpolation methods; the trained network results are read from the frozen study.
The learning curve uses training pools of 2, 4, 8, 12 and 16 subjects and the same six evaluation subjects throughout. Across the recorded study there were 15 TensorFlow fits, 90 evaluation cells and 7,830 slice predictions. Comparison with linear and cubic interpolation uses the corresponding withheld images. Relative error reduction compares baseline error with neural error, divided by baseline error; the tables report the study’s recorded median summaries.
At 16 training subjects, the recorded median masked-slice MAE reduction versus linear interpolation is 22.45%, and all six fixed evaluation cases improved. The evaluation panel comes from one collection and had been examined in earlier work, so it supports internal engineering development. Whole-acquisition completion and clinical validation remain separate tasks.
Reading the neural measurements
Compare like-for-like measurements within the same evaluation scope. The table explains their units and direction before the results.
| Measurement | Meaning / scale | Better direction |
|---|---|---|
| MAE | Mean absolute intensity error in HU; a percentage reduction is relative to the named baseline. | Lower error; higher reduction |
| SSIM | Structural similarity, unitless. Local masked measurements focus on hidden pixels; global measurements can include unchanged background. | Closer to 1 |
| Wasserstein distance | Difference between intensity distributions, in HU. Here it compares reconstructed and original slices; it does not require kVp groups. | Lower |
| Edge error | Difference in image gradients, in HU per pixel; examines anatomical boundary changes. | Lower |
Structural similarity does not imply quantitative HU fidelity. Read complementary measures together and inspect the full sample scope.
Measured reconstruction gains
The complete published experiment journey
Negative and plateaued results are part of the evidence, not hidden versions.