Machine Friendly Machine Learning: Interpretation of Computed Tomography Without Image Reconstruction

Hyunkwang Lee, Chao Huang, Sehyo Yune, Shahein Tajmir, Myeongchan Kim, Synho Do

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 applications 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 for sinogram-space machine learning: body region identification and intracranial hemorrhage (ICH) detection. Our proposed SinoNet 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 binary diagnosis testing, triage, and in clinical settings 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, Computed Tomography (CT) works by taking hundreds of X-ray projections from different angles. These raw measurements, called sinograms, are mathematically transformed into the 2D cross-sectional images that doctors study. However, this reconstruction step is fundamentally "human-centric." It is computationally expensive, time-consuming, and can discard or distort information present in the raw data. While humans find sinograms impossible to interpret—appearing as a blurry mess of overlapping waves—there is no inherent reason a machine should be limited to human-readable formats. By forcing AI to operate on reconstructed images, we may be introducing unnecessary latency and losing the "purity" of the original sensor data. This study explores "machine-friendly" machine learning by interpreting CT data directly in sinogram-space.

Methods

The researchers developed SinoNet, a customized Convolutional Neural Network (CNN) optimized for the unique geometry of sinograms. Unlike standard CNNs that use square filters for spatial features, SinoNet employs rectangular filters designed to capture features across the projection angle (horizontal) and detector position (vertical) axes.

The model was evaluated on two primary tasks:

  1. Body Part Recognition: Using 56,334 axial slices from 200 whole-body PET/CT exams, categorized into 16 regions.
  2. ICH Detection: Using 6,046 slices from 720 head CT scans, annotated by a consensus of neuroradiologists.

Since raw sensor data is rarely archived, sinograms were simulated by applying a 2D parallel-beam Radon transform to existing CT images. To simulate low-dose or rapid scans, the team tested the models on "sparsely sampled" data, reducing projections and detectors by factors of 3 and 9.

Comparison of reconstructed CT images and their corresponding sinograms for body parts and ICH detection.

Results

SinoNet achieved performance comparable to state-of-the-art models trained on reconstructed images. For body part recognition, it reached 96.6% accuracy on full sinograms, nearly matching the 97.4% of Inception-v3 on reconstructed images.

In ICH detection, SinoNet achieved an AUC of 0.918. Applying "windowing" techniques to the sinograms—analogous to how radiologists adjust contrast—improved performance to an AUC of 0.951. Crucially, in the most sparsely sampled scenario (40 projections), SinoNet maintained an AUC of 0.899, significantly outperforming image-space models which suffered from reconstruction artifacts in the low-data regime.

ROC curves for ICH detection across different sampling densities.

Discussion and Conclusion

The findings suggest that "human-interpretable" is not a requirement for "machine-actionable" diagnostic systems. SinoNet's superiority in the sparse regime indicates that reconstruction itself can be a bottleneck for low-dose imaging. This technology enables "edge-of-scanner" triage, where a scanner could prioritize a stroke patient for review immediately after data acquisition, before images are even rendered.

Limitations include the use of simulated sinograms, which do not perfectly capture real-world cone-beam geometries or Poisson noise. Future work must validate these models using raw measurement data directly from clinical scanners. While the "black box" nature of sinogram-space AI presents a challenge for clinical trust, its potential for rapid, low-dose diagnosis is significant.

Summary updated

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 applications 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 for sinogram-space machine learning: body region identification and intracranial hemorrhage (ICH) detection. Our proposed SinoNet 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 binary diagnosis testing, triage, and in clinical settings 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.
BibTeX
@inproceedings{lee2019machine,
  author = {Hyunkwang Lee and Chao Huang and Sehyo Yune and Shahein Tajmir and Myeongchan Kim and Synho Do},
  title = {{Machine Friendly Machine Learning: Interpretation of Computed Tomography Without Image Reconstruction}},
  booktitle = {Scientific},
  year = {2019},
  doi = {10.1038/s41598-019-51779-5},
}
Figure