Randomized Sparse Matrix Compression for Large-Scale Constrained Optimization in Cancer Radiotherapy

Authors: Shima Adeli, Mojtaba Tefagh, Gourav Jhanwar, Masoud Zarepisheh

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In this study, we demonstrate that using matrix sparsification techniques, primarily developed by the machine learning community, we can enjoy the computational efficiency of working with sparse matrices and still being able to solve constrained optimization problems within the clinical timeframe, without significantly compromising the integrity of the original problem that could potentially degrade the treatment quality. ... We demonstrate that applying a randomized sketch to the influence matrix A in radiotherapy optimization significantly outperforms the current naive sparsification approach. ... Our analysis utilized real-world data recently made publicly available through the opensource package Port Py [10]. We conducted experiments on data from 10 randomly selected lung patients...
Researcher Affiliation Collaboration Shima Adeli1 Mojtaba Tefagh1,2 Gourav Jhanwar3 Masoud Zarepisheh3 1Sharif University of Technology 2University of Edinburgh 3Memorial Sloan Kettering Cancer Center
Pseudocode Yes Algorithm 1 Randomized Minor-value Rectification (RMR) input A Rm n: The matrix to be sparsified, ϵ: The threshold for sparsification output S: The sparsified matrix S = deep copy of A for each row i in {1, 2, . . . , m} do Ti = {j | 0 < |aij| ϵ} set sij = 0 for all j Ti Σi = P j Ti |aij| ki = Σi/ϵ for t = 1, 2, . . . , ki do randomly select j Ti (with probability proportional to |aij|) and update sij sij + Σi/ki sign(aij) end for end for
Open Source Code Yes Code and sample data available at https://github.com/Port Py-Project/Compress RTP
Open Datasets Yes Our analysis utilized real-world data recently made publicly available through the opensource package Port Py [10].
Dataset Splits No The paper describes using full patient datasets for optimization and evaluation, but does not specify traditional train/validation/test splits for the dataset itself, as is common in supervised machine learning. Hyperparameters were varied and compared, but not via a distinct validation data split.
Hardware Specification Yes The experiments were conducted on a dual CPU system (Intel(R) Xeon(R) 6248 2.5GHz) running Windows 10 with 128 GB of RAM. ...we solved the original optimization problems on a powerful high-performance computing (HPC) system with approximately 320 GB of memory...
Software Dependencies Yes All optimization problems were modeled and solved using CVXPY [5] and MOSEK [13].
Experiment Setup Yes Each algorithm has a hyper-parameter threshold that determines the sparsity of the output matrix. To ensure a fair comparison, we ran each algorithm with various threshold values and compared the results based on different levels of sparsity in the output matrix (e.g., runtime of algorithm 1 vs. algorithm 2 for sparsity levels x, y). ...All optimization problems were modeled and solved using the penalty weights, ws +/ws , recommended by the Port Py package [10], along with the maximum and mean dose constraints presented in Table 3 in Appendix.