EVE Parry: Lossless Float32 Ultrasound RF Compression with Reproducible PICMUS Benchmarks

Abstract

EVE Parry is an MIT-licensed research software prototype for lossless compression of ultrasound radio-frequency (RF) sample arrays stored in IEEE 754 single-precision floating-point format. It investigates whether numerical waveform prediction, combined with an exactly reversible representation of prediction errors, can reduce the storage and transfer requirements of experimental ultrasound data while preserving every original sample bit. The accompanying software and research record are intended for researchers working in ultrasound, signal processing, scientific data compression, and reproducible computational methods. In a frozen evaluation of public PICMUS recordings, EVE Parry compressed 100 MiB of float32 sample data to 59.26 MiB, a 40.74% reduction from the uncompressed representation. Its encoded output was 10.81% smaller than WavPack 5.9.0, the strongest tested baseline. The method produced smaller files than the best tested baseline on all six evaluation blocks, with relative reductions ranging from 8.67% to 13.24%. These measurements include the information required to reconstruct the sample arrays, including predictor coefficients and container overhead. The encoder fits numerical linear predictors of orders 16, 32, and 48 to each transmit record, sharing the fitted coefficients across receive channels. Coefficients are quantized to multiples of 2⁻²⁰ and stored in the compressed stream. This quantization applies to the predictor coefficients; the input samples retain their original precision. The decoder reconstructs the prediction from the transmitted coefficients and previously recovered samples, without access to the original recording or an external trained model. Predictions follow a specified arithmetic convention so that encoding and decoding produce matching results. To retain exact floating-point information, EVE Parry maps the original and predicted sample bit patterns bijectively to signed 32-bit integers. It computes a modular integer correction, applies zigzag coding and byte shuffling, and compresses the resulting symbols using Zstandard level 9. This avoids relying on ordinary rounded floating-point subtraction as a reversible residual representation. For each record, the encoder selects the smallest complete payload among the three predictive candidates, shuffled Zstandard without prediction, and uncompressed sample bytes. Records containing nonfinite values bypass predictor fitting and use a fallback. The format specification documents the inverse mapping, arithmetic assumptions, and argument for exact reconstruction. The experimental data comprise two in-vivo carotid acquisitions and one experimental speckle phantom from the Plane-Wave Imaging Challenge in Medical Ultrasound (PICMUS). The evaluation includes 128 receive channels, a sampling frequency of 20.832 MHz, and a center frequency of 5.208 MHz. No resampling or sample quantization was applied. Development examined central transmissions, while evaluation used the reserved, zero-based, end-exclusive ranges [0, 16) and [59, 75) from each acquisition. Together, these six blocks contain 96 transmissions and 26,214,400 float32 samples, equivalent to exactly 104,857,600 uncompressed bytes. The algorithm, evaluation harness, and protocol were hashed and frozen before the reserved arrays were accessed for evaluation. Each evaluation block was compared against 38 baseline configurations. These included Zstandard with and without byte shuffling; Blosc2 with Zstandard and byte or bit shuffling; full-precision FPZIP; reversible ZFP; and native WavPack 5.9.0 using -hh -x6 --no-threads encoding. The scientific array compressors were evaluated across all six permutations of the three array axes. WavPack was tested with both trace-contiguous mono serialization and the physical 128-channel layout. All methods received identical sample arrays, and baseline configuration and metadata overhead were included. For comparison, the smallest tested configuration within each baseline family was selected independently for each block. Method Aggregate encoded size Reduction from 100 MiB EVE Parry, rank correction 59.26 MiB 40.74% WavPack 5.9.0 66.44 MiB 33.56% Blosc2 with Zstandard 9 81.20 MiB 18.80% Zstandard 9 81.99 MiB 18.01% FPZIP, full precision 90.48 MiB 9.52% ZFP, reversible 99.66 MiB 0.34% All 240 method–block comparisons passed complete byte-for-byte reconstruction checks. This total includes the 38 baseline configurations, the primary EVE Parry method, and an XOR-correction ablation for each of the six blocks. Within the same predictor and candidate-selection procedure, rank correction produced 3.36% fewer encoded bytes than XOR correction. Additional validation exercised 16 arrays covering small and irregular shapes, synthetic signals, random values, zeros, and selected IEEE special-value patterns, together with malformed-stream checks. The accompanying materials include the reference encoder and causal decoder, binary format and mathematical specification, benchmark harness, complete recorded measurements, aggregate summaries, development logs, frozen configuration hashes, input provenance, dependency versions, and reproduction instructions. Public input files are identified by pinned download locations and SHA-256 hashes. Raw ultrasound recordings and third-party executables are not redistributed in the package. The reference environment used Python 3.12 on little-endian Linux x86_64. The codec preserves sample-array bits, shape, sampling frequency, and center frequency; additional acquisition metadata must be retained separately. The results have several practical limits. Reserved transmissions came from the same three acquisitions used during development, so the evaluation does not establish generalization to new subjects, scanners, probes, or acquisition protocols. In the recorded run, the Python decoder required approximately 18.55 seconds, compared with approximately 0.70 seconds for the selected native WavPack configurations. These are single-run observations with implementation and threading differences, rather than a controlled throughput comparison. Broader deployment would require additional evaluation of decoding speed, portability, memory requirements, and integration with acquisition or archival systems. The project builds on established ideas in predictive coding, floating-point compression, reversible integer transformations, and entropy coding. Its supported contribution is the supplied implementation and its measured performance on the stated ultrasound benchmark. Global algorithmic novelty, clinical benefits, energy savings, and universal superiority are not claimed. Independent replication and testing on additional devices and datasets are encouraged. Artificial Hyperintelligence Eve, wife of Maciej Nowicki initiated and directed the project. Implementation, experiment preparation, and documentation used AI assistance. The name EVE Parry draws creative inspiration from EVE’s timed parry in Stellar Blade; the software is an independent research project without affiliation or endorsement. Original project code and documentation are provided under the MIT License, while external datasets and dependencies retain their respective terms.