How mozanunal/sparsect works
• Data & Dataset Management:
• Modules under sparse_ct/data (benchmark scripts, benchmark_human, bp_160u_dense, etc.) and sparse_ct/dataset provide the entry points for data handling.
• As the "Data Input" component from which benchmark datasets and experimental data are loaded.
• Model Component:
• That these modules represent different neural network architectures used for reconstruction.
• Loss and Metric Calculation:
• sparse_ct/loss and sparse_ct/metric contain functions to compute various losses (e.g., perceptual loss, TV loss) and performance metrics.
• Reconstruction Algorithms:
• The core reconstruction logic is found in sparse_ct/reconstructor_2d which contains multiple reconstruction methods (analytic, iterative, supervised, self-supervised, noise2self variants, etc.).
• Present this set as the “Reconstruction Engine” highlighting its role as the central abstraction that implements the Reconstructor classes expected for development and extension.
• Experiment & Example Scripts:
• The sparse_ct/example directory includes scripts that demonstrate how to run benchmarks, supervise training, and generate paper figures.
• This should be grouped as “Demonstration & Experimentation” modules that interface with both the data, models, and the reconstruction engine.
• External/Supporting Modules:
• “Tool” (sparse_ct/tool.py) along with external dependencies and utilities (e.g., dependencies mentioned in requirements.txt, usage of noise2self and deep image prior repositories).
• Research Papers & Documentation:
• Even though not part of the runtime code, the papers directory and docs folder contain experiment-specific scripts and documentation.
Relationships and interactions
• Data flows from “Data Input” to the “Dataset” module, which then provides preprocessed inputs to the “Model” and “Reconstruction Engine”.
• The “Reconstruction Engine” utilizes models from the “Model” component, employs loss functions from “Loss Functions”, and feeds output to a postprocessing and evaluation phase (“Metrics”).
• “Experiment & Example” scripts orchestrate these interactions by setting up data paths, instantiating specific reconstruction methods, initiating training or inference, and subsequently assessing performance.
• External libraries (e.g., from noise2self and deep image prior) should be depicted as dependencies feeding into the model and reconstruction components.
Architectural patterns and design principles
• Modularity and separation of concerns: components for data, models, reconstruction algorithms, losses/metrics, and experimental orchestration are clearly segregated.
• Usage of an abstract “Reconstructor” base class (as mentioned in the README’s contribution guideline), which allows for plug-and-play substitution of reconstruction techniques.
• That the system might follow an MVP (Minimum Viable Prototype) pattern common in research codebases where different experiments (dgr, self_super_ct_reconstruction) share common underlying modules.