Volumetric radio-frequency (RF) ultrasound texture analysis based on Nakagami statistics can quantify intra-tumor heterogeneity invisible in B-mode reading: Al-Kadi, Chung, Carlisle, Coussios and Noble (Medical Image Analysis 21(1):59–71, 2015) estimated 3D Nakagami shape/scale parametric maps from RF data and derived a multi-scale multifractal descriptor (MNF) that classified chemotherapy response of liver tumors with high accuracy on a single Zonare scanner. The authors explicitly flagged the decisive unresolved question: whether the learned statistical texture signature is a tissue property or partly an imaging-system artifact — i.e., whether a classifier trained on one ultrasound scanner can classify RF data from a different scanner without retraining — together with the coupled robustness questions of how steatosis-induced acoustic attenuation alters the Nakagami statistics and classification, how significant radiologist-label noise is and how to mitigate it, and whether the method can be used for staging liver tumors and longitudinally as an image-based biomarker of tumor growth and therapeutic response. These questions remain unaddressed for the uncalibrated Nakagami-fractal descriptor class; the problem remains open.

Background

When a diagnostic ultrasound pulse interrogates tissue, the envelope of the backscattered radio-frequency (RF) signal — the raw data before B-mode (the standard grayscale display image) processing — obeys statistics governed by the concentration and spatial arrangement of acoustic scatterers. The Nakagami distribution (P. Mohana Shankar, IEEE Trans. Ultrason., Ferroelectr., Freq. Control 47(3):727–736, 2000) is a two-parameter model of this envelope whose shape parameter mm indexes scattering regimes (from pre-Rayleigh, low scatterer density, to Rayleigh, dense random scatterers) and whose scale parameter tracks backscattered energy; mapping these parameters voxel-by-voxel yields Nakagami parametric images. Because a single estimation window resolves one spatial scale, and tumor texture is heterogeneous across scales, Al-Kadi et al. (Medical Image Analysis 21(1):59–71, 2015; DOI 10.1016/j.media.2014.12.004) proposed the MNF descriptor: Nakagami parameters are estimated at multiple window sizes over a voxel lattice, and local fractal (self-similarity) descriptors of the resulting multi-resolution parameter maps — which are invariant to affine intensity changes — are aggregated into a feature vector. Fractal dimension quantifies spatial self-similarity of the texture; multifractal analysis extends this to textures whose scaling varies locally, as in tumors. A naive Bayes classifier on MNF features separated progressive from non-progressive disease (referenced to RECIST, the Response Evaluation Criteria in Solid Tumors, scored on CT follow-up) with 98.95% accuracy on 227 RF cross-sections from 29 mouse hind-leg xenograft tumors (human tumor tissue implanted in mice; Zonare z.one, 10 MHz linear transducer) and 92.9% leave-one-out accuracy on 394 clinical liver-tumor RF cross-sections acquired before chemotherapy (Zonare z.one, 4 MHz curvilinear transducer). All data came from a single scanner model at one institution, with training labels from the consensus of two radiologists. The published discussion states the resulting limitations as future work: (i) training samples may be mislabelled, and the significance of this and mitigation strategies are unexplored; (ii) hepatic steatosis (fatty liver) attenuates tissue properties and it is unknown how steatosis affects MNF classification accuracy; (iii) RF characteristics and hence speckle appearance differ between ultrasound devices, and it is unknown whether a training set from one scanner can classify images from a different scanner or the results are scanner-specific. The conclusion adds that staging of liver tumors and longitudinal use of the method as an image-based biomarker of tumor growth and therapeutic response remain to be investigated. Independent lines of work make this question sharp rather than hypothetical. Larrue and Noble (Ultrasound in Medicine and Biology 40(5):917–930, 2014) showed that Nakagami imaging is sensitive to estimation errors from attenuation and beamformation choices. Christensen, Rosado-Mendez and Hall (Ultrasound in Medicine and Biology 50(12):1800–1811, 2024) showed that depth-dependent power loss — a system- and processing-dependent effect — biases speckle-statistics estimates, a concrete mechanism by which the uncalibrated envelope statistics used by MNF could carry a system signature. Conversely, calibrated quantitative ultrasound (QUS), in which the backscatter and attenuation coefficients are recovered against a reference phantom (Zagzebski, Rosado-Mendez, Nasief and Hall, AIP Conf. Proc. 1747:050001, 2016), has demonstrated a degree of cross-platform portability: in layered tissue-mimicking phantoms imaged by three clinical scanners (Nam et al., Ultrasonic Imaging 34(4):209–221, 2012) and in an in vivo rodent fibroadenoma model across four imaging systems (Wirtzfeld et al., IEEE Trans. Ultrason., Ferroelectr., Freq. Control 60(7):1386–1400, 2013); Zaidman, Holland and Hughes (Ultrasound in Medicine and Biology 38(9):1618–1625, 2012) for skeletal muscle; Han et al. (European Radiology 29(9):4699–4708, 2019) for liver attenuation/backscatter in patients with NAFLD (non-alcoholic fatty liver disease) across two manufacturer platforms; and Sannachi et al. (Ultrasound in Medicine and Biology 46(5):1142–1157, 2020), who found good agreement of QUS-texture breast-tumor response detection across two clinical scanners, with tissue heterogeneity dominating over scanner beam properties. Whether such portability extends to uncalibrated Nakagami-fractal descriptors for liver tumor response — and whether the confounding strands (steatosis, label noise) permit a deployable biomarker — is the open question. On the steatosis side, envelope-statistics imaging is itself used to detect hepatic fat (e.g., Lin et al., Ultrasonics 142:107391, 2024), confirming that steatosis shifts precisely the statistics MNF relies on, which substantiates but does not answer the confounding question.

Problem Statement

Determine whether volumetric Nakagami-statistics texture characterization of tumors, as instantiated by the MNF descriptor of Al-Kadi et al. (2015) — a multi-scale multifractal feature vector computed from Nakagami shape/scale parametric maps of RF backscatter envelopes — encodes a tissue-specific signature that generalizes across ultrasound imaging systems and realistic clinical confounders. Concretely, on the source's own future-work agenda: (1) the central cross-scanner question — determine whether the descriptor's discriminative signal for chemotherapy-response classification is attributable to tissue or to scanner-specific RF chain, beamforming, and transducer characteristics, by training a classifier on RF data from one ultrasound system and evaluating, without retraining, on RF data from at least one other system (with either outcome — successful transfer or demonstrated scanner-specificity — being a valid answer); (2) the coupled confounder question — measure how hepatic steatosis-induced acoustic attenuation alters the Nakagami shape/scale statistics and fractal signatures and whether it degrades response classification; (3) the training-label question — quantify the sensitivity of the classification to radiologist-consensus label noise and demonstrate mitigation strategies; and (4) investigate whether the descriptor can be used for staging liver tumors and, longitudinally, as an image-based biomarker of tumor growth and therapeutic-response monitoring. The central cross-scanner component (1) is the core of the question; components (2)–(4) are separately resolvable robustness questions bearing on the same characterization. These components jointly answer one scientific question: is the hidden heterogeneity signature a property of the tissue that RF acquisition systems can recover, or an artifact-bound property of the specific acquisition?

References

  1. Al-Kadi OS, Chung DYF, Carlisle RC, Coussios CC, Noble JA (2015). Quantification of ultrasonic texture intra-heterogeneity via volumetric stochastic modeling for tissue characterization. Medical Image Analysis 21(1):59–71. DOI: 10.1016/j.media.2014.12.004; also arXiv:1601.03531. https://doi.org/10.1016/j.media.2014.12.004
  2. Shankar PM (2000). A general statistical model for ultrasonic backscattering from tissues. IEEE Transactions on Ultrasonics, Ferroelectrics, and Frequency Control 47(3):727–736. DOI: 10.1109/58.842062. https://doi.org/10.1109/58.842062
  3. Larrue A, Noble JA (2014). Modeling of errors in Nakagami imaging: illustration on breast mass characterization. Ultrasound in Medicine and Biology 40(5):917–930. DOI: 10.1016/j.ultrasmedbio.2013.11.018. https://doi.org/10.1016/j.ultrasmedbio.2013.11.018
  4. Al-Kadi OS, Chung DYF, Coussios CC, Noble JA (2016). Heterogeneous tissue characterization using ultrasound: a comparison of fractal analysis backscatter models on liver tumors. Ultrasound in Medicine and Biology 42(7):1612–1626. DOI: 10.1016/j.ultrasmedbio.2016.02.007; also arXiv:1912.09903. https://doi.org/10.1016/j.ultrasmedbio.2016.02.007
  5. Al-Kadi OS (2016). Multiscale Nakagami parametric imaging for improved liver tumor localization. IEEE International Conference on Image Processing (ICIP) 2016, pp. 3384–3388. arXiv:1906.04333. https://arxiv.org/abs/1906.04333
  6. Al-Kadi OS, Van De Ville D, Depeursinge A (2016). Multidimensional Texture Analysis for Improved Prediction of Ultrasound Liver Tumor Response to Chemotherapy Treatment. MICCAI 2016, Lecture Notes in Computer Science, pp. 619–626. DOI: 10.1007/978-3-319-46720-7_72. https://doi.org/10.1007/978-3-319-46720-7_72
  7. Nam K, Rosado-Mendez IM, Wirtzfeld LA, Ghoshal G, Pawlicki AD, Madsen EL, Lavarello RJ, Oelze ML, Zagzebski JA, O'Brien WD, Hall TJ (2012). Comparison of ultrasound attenuation and backscatter estimates in layered tissue-mimicking phantoms among three clinical scanners. Ultrasonic Imaging 34(4):209–221. DOI: 10.1177/0161734612464451. https://doi.org/10.1177/0161734612464451
  8. Wirtzfeld LA, Nam K, Labyed Y, Ghoshal G, Haak A, Sen-Gupta E, He Z, Hirtz NR, Miller RJ, Sarwate S, Simpson DG, Zagzebski JA, Bigelow TA, Oelze M, Hall TJ, O'Brien WD (2013). Techniques and evaluation from a cross-platform imaging comparison of quantitative ultrasound parameters in an in vivo rodent fibroadenoma model. IEEE Transactions on Ultrasonics, Ferroelectr, and Frequency Control 60(7):1386–1400. DOI: 10.1109/TUFFC.2013.2711. https://doi.org/10.1109/TUFFC.2013.2711
  9. Zaidman CM, Holland MR, Hughes MS (2012). Quantitative Ultrasound of Skeletal Muscle: Reliable Measurements of Calibrated Muscle Backscatter from Different Diagnostic Ultrasound Systems. Ultrasound in Medicine and Biology 38(9):1618–1625. DOI: 10.1016/j.ultrasmedbio.2012.04.020. https://doi.org/10.1016/j.ultrasmedbio.2012.04.020
  10. Zagzebski JA, Rosado-Mendez IM, Nasief HG, Hall TJ (2016). Quantitative ultrasound: Enhancing diagnosis using estimates of acoustic attenuation and backscatter. AIP Conference Proceedings 1747:050001. DOI: 10.1063/1.4954108. https://doi.org/10.1063/1.4954108
  11. Han A, Zhang YN, Boehringer AS, Andre MP, Erdman JW, Loomba R, Sirlin CB, O'Brien WD (2019). Inter-platform reproducibility of ultrasonic attenuation and backscatter coefficients in assessing NAFLD. European Radiology 29(9):4699–4708. DOI: 10.1007/s00330-019-06035-9. https://doi.org/10.1007/s00330-019-06035-9
  12. Sannachi L, Gangeh M, Naini AS, Bhargava P, Jain A, Tran WT, Czarnota GJ (2020). Quantitative Ultrasound Monitoring of Breast Tumour Response to Neoadjuvant Chemotherapy: Comparison of Results Among Clinical Scanners. Ultrasound in Medicine and Biology 46(5):1142–1157. DOI: 10.1016/j.ultrasmedbio.2020.01.022. https://doi.org/10.1016/j.ultrasmedbio.2020.01.022
  13. Osapoetra LO, Sannachi L, Quiaoit K, Dasgupta A, DiCenzo D, Fatima K, Wright F, Dinniwell R, Trudeau M, Gandhi S, Tran W, Kolios MC, Yang W, Czarnota GJ (2021). A priori prediction of response in multicentre locally advanced breast cancer (LABC) patients using quantitative ultrasound and derivative texture methods. Oncotarget 12(2):81–94. DOI: 10.18632/oncotarget.27867. https://doi.org/10.18632/oncotarget.27867
  14. Lin CH, Ho MC, Lee PC, Yang PJ, Jeng YM, Tsai JH, Chen CN, Chen A (2024). Clinical performance of ultrasonic backscatter parametric and nonparametric statistics in detecting early hepatic steatosis. Ultrasonics 142:107391. DOI: 10.1016/j.ultras.2024.107391. https://doi.org/10.1016/j.ultras.2024.107391
  15. Christensen A, Rosado-Mendez IM, Hall TJ (2024). A Study on the Effects of Depth-Dependent Power Loss on Speckle Statistics Estimation. Ultrasound in Medicine and Biology 50(12):1800–1811. DOI: 10.1016/j.ultrasmedbio.2024.08.001. https://doi.org/10.1016/j.ultrasmedbio.2024.08.001
  16. Varghese BA, Rivas M, Cen S, Lei X, Chang M, Lee K, Jamie J, Amoedo R, Franco M, Hwang D, Desai B, King K, Cheng P, Duddalwar V (2023). Tissue characterization of renal masses using Nakagami-modeling of ultrasound-based texture. SPIE Medical Imaging 2023 (Proceedings of the SPIE, volume 12468). DOI: 10.1117/12.2670273. https://doi.org/10.1117/12.2670273
  17. Zhang X, Zhang Q, Zhang T, Song G, Wang P (2026). Fractal-guided multi-scale contrastive learning for robust liver tumor classification in ultrasound. Engineering Applications of Artificial Intelligence 178:115047. DOI: 10.1016/j.engappai.2026.115047. https://doi.org/10.1016/j.engappai.2026.115047
  18. Hall TJ, Fowlkes JB, Zahlmann G (2025). International standardization efforts in elastography, quantitative ultrasound and beyond. Ultrasound in Medicine and Biology 51(Suppl. 1):S49–S50 (WFUMB 2025 congress abstract). DOI: 10.1016/j.ultrasmedbio.2025.10.172. https://doi.org/10.1016/j.ultrasmedbio.2025.10.172