Ultrasound Signal Processing

& Artificial Intelligence in Medicine

Pradeep Chaudhary, PhD

Research Fellow

Mayo Clinic, USA

Content

Ultrasound Signal Processing & AI in Medicine

​

​

​

• What is ultrasound ?

​

• Vibration analysis Via Ultrasound

​

• Ultrasound-based Urodynamic Study for Bladder Pressure Estimation

​

• Ultrasound Localized Microscopy

​

• B-mode + Signal processing + CNN

​

​

​

​

​

​

What is ultrasound?

What is sound?

High-level messages

Ultrasound Signal Processing & AI in Medicine

What is Ultrasound?

Ultrasound Signal Processing & AI in Medicine

Adopted from community element14.com

Significance of Ultrasound

Ultrasound Signal Processing & AI in Medicine

Why Ultrasound

• Non-invasive
• Real time
• Portable
• Convenient

​

Uses of Medical US

• Diagnostics
• Medical procedures
• Therapeutics

​

Ultrasound Technique

Ultrasound Signal Processing & AI in Medicine

https://www.genosound.com/

Imaging modalities

Ultrasound Signal Processing & AI in Medicine

Research ultrasound device

(verasonics)

Vibration analysis

Via Ultrasound

Why monitor bone health in preterm infants?

​

Ultrasound Signal Processing & AI in Medicine

• Metabolic Bone Disease of Prematurity (MBDP)

• MBDP remains common despite improved neonatal care; early detection helps guide nutrition/supplementation and reduce fracture risk.

​

• Existing tools are challenging in infants: DEXA has practical limitations.

​

• Conventional QUS can be affected by soft tissue path and positioning.

​

• Goal: a bedside, repeatable, low‑risk method for longitudinal monitoring.

Experimental setup

Acquisition

• Ultrasound ARF push induces broadband vibration in tibia.
• Hydrophone placed on skin records resulting acoustic signal.
• Repeat at multiple focal locations to improve reliability.

Received hydrophone data

Wave parameters used as bone biomarkers

• Speed‑of‑Sound (SoS) for fast and slow waves (SoS1, SoS2).
• Mean Peak Frequency (MPF) from time‑frequency energy (MPF1, MPF2).
• Track change with postmenstrual age and weight (longitudinal visits).

Quality → Denoise → Segment fast/slow waves → Estimate SoS (time‑delay) + MPF (STFT)

Received hydrophone data

Biomarkers

1) Speed‑of‑Sound (SoS)

• Record signals at multiple ARF focal locations (F1…F4) separated by Δd along the bone surface.
• Estimate inter‑location time delay (Δt) using cross‑correlation (after denoising + segmentation).
• Compute SoS = Δd / Δt for fast wave (SoS1) and slow wave (SoS2).

2) Mean Peak Frequency (MPF)

• Compute STFT of the segmented fast and slow wave portions.
• At each time, find the frequency with maximum energy; average over time → MPF1/MPF2.
• Use MPF as a frequency‑domain marker of the vibrational response.

• 35 infants enrolled (10 full term, 25 preterm); data collected longitudinally for preterm infants (up to 4 visits).
• Ultrasound: Verasonics Vantage 128 with L11‑5v; ARF push ~7.8 MHz (5‑cycle tone burst); hydrophone records.
• Measurements taken on tibia; both legs when feasible; quality control removes noisy segments/locations.

Outputs analyzed: SoS1, SoS2, MPF1, MPF2 vs postmenstrual age (weeks) and weight (kg)

Invivo Dataset

• Bilateral agreement (left vs right): Bland‑Altman analysis + ICC with 95% CI.🡪 measure how two leg measurements are correlated.

​

​

Statistical analysis

• Association models: linear mixed‑effects models linking wave parameters to postmenstrual age and weight.

​

Statistical analysis

Take aways

• Higher SoS generally indicates stiffer / more mineralized bone.

SoS1/SoS2 show significant positive association with postmenstrual age and with weight.

​

• MPF shifts can reflect changes in the vibrational spectrum over maturation.

MPF1/MPF2 also increase with maturation (age/weight) with statistically significant effects in the pilot cohort.

​

• Longitudinal trends are more important than single‑timepoint absolute values.

Next steps

• Larger cohorts + outcome linkage (DEXA/QUS/clinical endpoints) to validate biomarkers.
• Improve standardization (probe placement, coupling, motion artifact handling).
• Explore ML/AI on multi‑parameter features for risk stratification and automated QC.

Ultrasound-based Urodynamic Study for Bladder Pressure Estimation

Motivation

Why replace catheter-based UDS?

Goal: non-invasive ultrasound-only framework to estimate Pves, Pabd, and Pdet and derive clinical indices.

UDS invasive: UDS is the clinical gold standard for LUT dysfunction diagnosis, but it is invasive (urethral + rectal catheters).

​

Discomfort/embarrassment and complications (pain, bleeding, infection) reduce adoption and repeatability

​

Artifacts/ errors: Body movement, pump, less doctor are trained for UDS analysis

Need real-time

Ultrasound acquisition during UDS

In vivo acquisition (summary)

• System: Verasonics Vantage + curved array transducer (C5).
• Ultrafast imaging: 5 plane-wave angles (−4° to 4°), PRF ≈ 4.717 kHz; beamformed IQ exported.
• US data acquired every ~6 seconds during the UDS procedure; sync marks (cough/Valsalva) used for time alignment.

Signals available during UDS

• Pves, Pabd (catheter-derived)
• EMG
• Flow
• Filled volume
• Computed: Pdet = Pves − Pabd

Overall pipeline

. . .

High frame rate IQ data

[410 ×365×55]

Bladder wall segmentation

​

US imaging acquisition during UDS every 6 seconds

EMG

Flow

Filled volume

. . .

Along bladder wall

Along anterior tissue

Displacement plots

• Bladder thickness
• Bladder wall polynomial coefficient

Autocorrelation technique for Particle velocities

temporal filtering (moving average of three frames)

Wave tracking (displacement plot)

Proposed CNN model

1) Direct Pdet prediction from displacement plots + bladder features (ResNet‑50 regression).

​

2) Indirect Pdet estimate: Pdet = Pves − Pabd, where Pves and Pabd are predicted by separate models.

​

3) Final refinement: concatenate (Pdet_direct, Pdet_indirect) and pass through a small regressor:

• FC(32) → FC(16) → Regression (final Pdet).

Motivation: mitigate artifacts/inconsistencies in UDS-derived ground truth and improve stability of peaks

Training & validation

Data and preprocessing

Subjects: 27 (majority male) with bladder overactivity cohort (as reported).

​

Remove UDS samples with negative pressures (Pves/Pabd/Pdet) and segments with poor wave propagation (noisy displacement plots).

​

Time alignment between UDS and US via sync/maneuver marks.

Prediction Results

Pdet > 15cmH20 consider alarm

Clinical interpretability: BOOI & BCI

Indices derived from predicted pressures/flow

BOOI = Pdet@Qmax − 2·Qmax

BCI = Pdet@Qmax + 5·Qmax

​

Clinical cutoffs:

• BOOI > 40 obstructed; 20–40 equivocal; <20 unobstructed

​

• BCI >150 strong; 100–150 normal; <100 weak

Take aways

US‑UDS: non‑invasive pressure estimation + clinically interpretable outputs

Ultrasound wave dynamics (displacement plots) + bladder anatomy features can predict Pves, Pabd, and Pdet.

​

Transfer learning (modified ResNet‑50) + multi-stage refinement improves Pdet robustness.

High event-detection AUC and strong agreement for BOOI/BCI suggest real clinical utility.

​

Limitations & future work

• Dataset size is currently limited; expanding patient cohort should improve robustness and generalization.

​

• Ground truth UDS contains artifacts; better artifact labeling/handling could further stabilize training.

​

• Automate manual steps (wall segmentation, thickness, volume) to enable a fully ultrasound-only pipeline.

​

• Explore training a dedicated network from scratch (domain-specific) once sufficient data is available.

​

​

Ultrasound Localized Microscopy

Motivation behind using orientation biomarker

Fig: Stage of angiogenesis

Small avascular tumor

Tumor secretion of angiogenesis factor

Neovascularization

(orientation pattern)

Rapid growth of tumor and the microvessels

• Low oxygen levels in early-evolving tumors trigger the secretion of vascular endothelial growth factors, promoting neovascularization and tumor progression.

​

• The newly formed microvessels in malignant tumors are typically leaky, tortuous, irregular, and often oriented toward the center of the lesion.

​

• Qualitative analysis and quantitative biomarker can help distinguishing malignant cases from benign.

​

• Quantitative biomarkers based on morphology is well explored in the literature (Stage 3 and 4 mostly).

​

• Quantitative analysis based on orientation is still missing in the literature (Stage 3).

​

​

​

Stage 1:

Stage 2:

Stage 3:

Stage 4:

HDMI Technique

Fig: Steps for HDMI formation

Proposed orientation biomarkers

• Penetrating microvessels appear like the vertical line.

• Circumferential microvessels appear like a horizontal line.

Polar coordinate-based biomarkers

Ѳ

r

Radial histogram

Tangential histogram

1

2

3

4

Proposed orientation biomarkers

Polar coordinate-based biomarkers (Contd…)

PGA 🡪 180

PGA 🡪 0

Polar gradient angle (PGA): Provide directionality of microvessel. Gradient operation is used to provide directionality of change of pixel

Fig: Polar gradient angle (PGA)

5

Fig: Angle based penetration density (APD)

Cartesian coordinate-based biomarkers

6

Proposed orientation biomarkers

Fig: PCD biomarker relative distance measurement.

• This study included a total of 70 subjects, each having at least a single ultrasound-identifiable breast lesion, mostly classified as BI-RADS 4 and up, and recommended for core needle breast biopsy.
• An Alpinion ultrasound system Ecube12-R (ALPINION Medical Systems, Seoul, South Korea), equipped with a linear array running at 8.5MHz, L3-12H (ALPINION Medical Systems) was used to acquire the images. First, plane-wave imaging mode was employed to identify the breast mass in the B-mode image. Subsequently, a series of high-frame-rate images at 600 frames per second were captured at each tumor site, where each frame was formed using 5-angle coherent plane-wave compounding.

​

Database

Benign

Malignant

Fig: (A, D) B-mode images with lesion boundary (blue) and 2 mm dilation (green).� (B, E) HDMI images of the corresponding lesions.� (C, F) HDMI images after conversion into polar coordinates.

Results

• Univariable and multivariable analysis is performed.
• For multivariable analysis principal component analysis was used. First 3 principal component is used for classification task.
• For univariable analysis threshold-based classifier is used. Optimal cutpoint was using the Youden index.
• For multivariate analysis linear regression model was used
• Proposed multivariate model and combining it with BIRADS give good classification performance.

Table : Univariable analysis

Results

• Orientation-based biomarkers derived from microvessel patterns provide quantitative insight into tumor vascular architecture.
• Malignant lesions show penetrating, feeder-type vasculature, whereas benign lesions exhibit circumferential vessel organization.
• Combining orientation biomarkers with morphological features improves classification accuracy and may help reduce unnecessary biopsies.
• Future work 1: integration of 3D orientation analysis and validation on larger clinical datasets to enhance diagnostic precision.
• Future work 2: scoring the lesion based on orientation biomarker threshold to reduce feature dimensionality.
• Future work 3: Analyze the performance for small size and small number of microvessel. Can help in early diagnosis.

​

​

​

Results

B-mode + Signal processing + CNN

Motivation

Why ultrasound CAD for breast cancer?

• Need timely diagnosis to reduce mortality.
• Mammograms can be noisy and lesions may be subtle; overlapping anatomy can obscure findings.
• Ultrasound is relatively low-risk and low-cost, but interpretation can vary—motivating computer-aided diagnosis (CAD).
• Goal: robust, accurate, less human-dependent classification of benign vs malignant (and normal where available).

Datasets

Two public breast ultrasound datasets used for evaluation

BUSI (Breast Ultrasound Images)

• 780 images (2018; Baheya Hospital, Cairo, Egypt)
• 3 classes: normal 133, malignant 210, benign 487
• Average image size ~500×500

Dataset-B (Breast Ultrasound Lesion Dataset)

• 163 images (2012; UDIAT, Sabadell, Spain)
• 2 classes: benign 110, malignant 53
• Mean size ~760×570

Sample images (paper Fig. 1)

Key point: Dataset-B is small → motivates transfer learning + robust feature mapping

Main contributions

What is new here?

• Iterative Fourier–Bessel Decomposition Method (IFBDM) 2D signals uses Fourier–Bessel order-domain boundaries to guide iterative filtering and extract meaningful modes/sub-bands.
• Multi-resolution breast ultrasound analysis: decompose each image into informative sub-band images (detail + approximation) before classification.
• Pre-trained CNN + kernel-based feature mapping: extract deep features via transfer learning, then apply kernel functions (e.g., RBF, polynomial, linear) to improve feature space and robustness for small datasets.

Proposed framework (overview)

End-to-end pipeline used in the paper

IFBDM: what the decomposition produces

Sub-bands provide multi-resolution views of texture/edges vs smooth structure

High-level intuition

• IFBDM splits an image into sub-band components (modes).
• One sub-band emphasizes high-frequency detail (edges/texture).
• Another sub-band captures lower-frequency structure (approximation/smooth components).
• Combining information across sub-bands can outperform using the raw image alone.

Example sub-bands (paper Fig. 4)

CNN-kernel features: why add kernels on top of deep features?

Step 1: Transfer learning (feature extraction)

• Use a pre-trained CNN backbone (e.g., ResNet-50).
• Freeze most layers; replace/fine-tune the final layer for the dataset classes.
• Extract deep features from the final trainable layer (per sub-band).

Step 2: Kernel mapping (feature space enrichment)

• Apply kernel functions to deep features: RBF, polynomial, or linear.
• Goal: improve separability/robustness, especially with limited data.
• Then ensemble features across sub-bands and classify (SVM / Random Forest).

IFBDM

sub-bands

Pre-trained

CNN

Deep features (per sub-band)

Kernel mapping (RBF / POL / Linear)

Feature

ensemble

Classifier

(SVM / RF)

Output: benign / malignant

(+ normal in BUSI)

Key idea: kernels can make deep features more discriminative with limited training data

Which CNN backbone works best?

Performance comparison across pre-trained CNNs (Perf_avg)

• ResNet-50 ranks highest on average performance for both BUSI and Dataset-B in the paper’s comparison.
• use a strong backbone for feature extraction, then focus on kernels + classical classifiers.

Perf_avg = (Accuracy + Sensitivity + Specificity) / 3 (as defined in the paper)

What helps most? Kernels and decomposition (ablation)

Table II: kernel features vs no kernel, and IFBDM vs no decomposition

Key observations

• Polynomial (POL) kernel consistently improves over “without kernel”.
• Using IFBDM sub-bands improves over “without decomposition”.
• Best accuracy reported: BUSI 98.68% and Dataset-B 98.13% (POL + RF).

Highlighted: best performance

Comparison with prior methods (state-of-the-art)

Reported best metrics (paper Table III)

The proposed method is shown at the bottom of each dataset block and reports the best overall performance in this table.

How to read this

• BUSI (3-class available): accuracy 98.68% with strong SEN/SPE and AUC ~0.99.
• Dataset-B (2-class): accuracy 98.13%, high SEN/SPE and AUC ~0.998.
• The gain is attributed to multi-resolution sub-bands + kernel-mapped deep features.

Conclusions

Main conclusions (as stated in the paper)

• IFBDM provides meaningful multi-resolution sub-bands and improves classification vs using the raw image alone.
• Kernel-based mapping on top of deep features can improve separability/robustness.
• Best results in the paper use ResNet-50 deep features + polynomial kernel + Random Forest.

Limitations / next steps (practical)

• Dataset sizes are limited (especially Dataset-B) → risk of optimistic estimates without careful cross-validation.
• External validation + standardized acquisition/labeling are required for clinical translation.

Q & A

Thank you