Abstract
Recent advancements in deep learning for automated image processing and classification have accelerated many new applications for medical image analysis. However, most deep learning algorithms have been developed using reconstructed, human-interpretable medical images. While image reconstruction from raw sensor data is required for the creation of medical images, the reconstruction process only uses a partial representation of all the data acquired. Here, we report the development of a system to directly process raw computed tomography (CT) data in sinogram-space, bypassing the intermediary step of image reconstruction. Two classification tasks were evaluated for their feasibility of sinogram-space machine learning: body region identification and intracranial hemorrhage (ICH) detection. Our proposed SinoNet, a convolutional neural network optimized for interpreting sinograms, performed favorably compared to conventional reconstructed image-space-based systems for both tasks, regardless of scanning geometries in terms of projections or detectors. Further, SinoNet performed significantly better when using sparsely sampled sinograms than conventional networks operating in image-space. As a result, sinogram-space algorithms could be used in field settings for triage (presence of ICH), especially where low radiation dose is desired. These findings also demonstrate another strength of deep learning where it can analyze and interpret sinograms that are virtually impossible for human experts.
Introduction
In traditional radiology, CT scanners collect raw sensor data which is then converted into a 2D or 3D image through complex mathematical reconstruction, such as Filtered Back Projection. This process is inherently "human-centric," existing primarily because radiologists require visual representations to make diagnoses. However, reconstruction is computationally expensive, time-consuming, and can introduce artifacts or lose subtle information present in the raw data.
Most medical AI models are trained on these reconstructed images. This study addresses a fundamental question: If the goal is automated diagnosis, why bother with the reconstruction step at all? By training models to "read" raw, non-interpretable sinograms, we can achieve faster triage and enable the use of significantly lower radiation doses.
Methodology
The researchers developed SinoNet, a customized Convolutional Neural Network (CNN) based on the Inception architecture. A key innovation in SinoNet is the use of rectangular filters instead of standard square filters. This design choice accounts for the specific geometry of sinograms, where one axis represents the projection angle and the other represents the detector position, allowing the model to better capture inherent sinusoidal patterns.
The performance of SinoNet was compared against a standard Inception-v3 model trained on reconstructed CT images across three levels of data quality: fully sampled, moderately sampled (1/3 data), and sparsely sampled (1/9 data). The study utilized two primary datasets: body part recognition (56,334 axial slices) and intracranial hemorrhage (ICH) detection (720 head CT scans).
Results: Body Part Recognition
For the task of body region identification, both sinogram-space and image-space models performed exceptionally well. Accuracies ranged from 96.2% to 97.9% across all sampling densities. This indicates that the raw data contains sufficient spatial information for broad anatomical classification, even when significantly downsampled.
Results: Intracranial Hemorrhage (ICH) Detection
The most significant advantages of sinogram-space learning were observed in ICH detection. On fully sampled data, SinoNet achieved an AUC of 0.918, compared to 0.898 for the image-space model.
The performance gap widened in the sparsely sampled domain, which simulates low-dose CT scans. While the image-based model's performance dropped to an AUC of 0.852 due to reconstruction noise, SinoNet maintained a robust AUC of 0.899. Furthermore, applying "windowing" (adjusting contrast levels) to the input data consistently improved performance for both modalities.
Discussion and Conclusion
The study proves that deep learning can extract diagnostic features from data that is virtually impossible for a human expert to interpret. The superiority of SinoNet in the sparse domain suggests that diagnostic-grade triage could be performed using up to 89% less radiation than standard CT scans.
By skipping the reconstruction step, AI could be embedded directly into "smart scanners" or portable field devices for near-instantaneous triage in emergency settings. Future work will involve validating SinoNet on actual raw data from clinical scanners to account for real-world noise patterns and complex geometries.
Summary updated