Medical Applications of AI

“Days of Miracle and Wonder”

- Paul Simon

Three Examples

• Quantum applications - MEG and the ghost camera
• Plan AI - Program of radiation - straight lines vs. biology. (Linear regression vs. the world.) A way to figure out the topology of the body.
• The magic stethoscope - a mature AI technology. (Is it still learning ?)

Quantum ghost camera

Entagled photons of different Ultra-low light imaging: By using heralded single photons, background counts can be virtually eliminated, achieving high-quality images from fewer than one detected photon per image pixel on average News-Medical

Why do such a strange thing ?

Red light cameras are. Great for biologic specimens

Blue light camera are cheap, well-researched

Medical imaging with reduced radiation dose

Imaging through turbulent or obscuring media

Low-light imaging scenarios

Remote sensing and surveillance

How good is AI ?

.3 microns

1/20 the diameter of a human hair

How does AI do it ?

A combination of quantum and digital computing.

Planning radiation therapy

• Why does radiation work ? Differential rate of replication from surrounding normal cells
• Radiation tx is complicated - many reasons
• No straight lines in biology
• Contouring
• Tumor has white cells, blood vessels, and connective tissue within it.
• Need good models of the surrounding biologic substrate of the organ involved.

U-net and variants

Basic Architecture

U-Net has a distinctive "U" shape when you diagram it, which is where it gets its name. It consists of two main pathways:

1. Contracting Path (Encoder): The left side of the "U" that progressively downsamples the input image, capturing context and high-level features. It uses convolutional layers and pooling operations to compress the spatial dimensions while increasing the number of feature channels.
2. Expanding Path (Decoder): The right side of the "U" that progressively upsamples the features back to the original image resolution. This enables precise localization and pixel-level predictions.

The Key Innovation: Skip Connections

What makes U-Net special are the horizontal "skip connections" that bridge across the U, connecting layers from the encoder directly to corresponding layers in the decoder. These connections allow the network to combine:

• High-resolution spatial details from early layers (where is something?)
• Abstract semantic information from deep layers (what is it?)

This is crucial for medical image segmentation because you need both precise boundaries and contextual understanding.

Cardiac fun facts

• Every heart is different
• Conduction system of the heart are modified muscle cells
• Daily — 100,000 beats. Lifetime 2.5-3.0 billion beats
• Contracts on its own
• Innervation - post-transplant
• Power of the electrical energy of the heart is a bit less than a small LED -batteries to self-power pacemakers are being developed

The Heart is a Pump

“All we know is still infinitely less than all that unknown.”

- William Harvey

Key Aspects of Harvey's Model:

Heart as a Pump: Harvey established that the heart (specifically the left ventricle) acts as a pump,

contracting actively to expel blood into the arteries, with the pulse being the result of this action.

Continuous Circuit: Blood moves in a continuous circuit, rather than being used up by tissues.

Directional Flow and Valves: He demonstrated, using ligatures, that venous blood flows only toward

the heart, aided by valves that prevent backflow.

Quantitative Proof: By calculating the volume of blood pumped by the heart, Harvey showed that it was

impossible for the liver to create such a massive amount of blood daily, proving it must be recycled.

Connection of Vessels: While he did not see them, Harvey predicted the existence of microscopic vessels

(capillaries) connecting arteries to veins. 

1578 – 1657

The magic stethoscope

Active Noise Cancellation (ANC): Reduces, background noise in busy environments (ED, ICU),

allowing for better focus on body sounds.

Sound Amplification:.

Visualization & Recording: The ability to visualize phonocardiograms (heart sound waveforms)

on a smartphone or computer screen.

​

​

​

​

With AI enhancement can immediately identify:

​

1. Murmurs - valvular problem

​

2. Heart Failure - with and without preserved ejection fraction

​

3. Atrial Fibrillation

​

4. Pulmonary hypertension

​

SYSTEMATIC REVIEW article

Front. Artif. Intell., 12 November 2024

Sec. Machine Learning and Artificial Intelligence

Volume 7 - 2024 | https://doi.org/10.3389/frai.2024.1434022

A review on deep learning methods for heart sound signal analysis

Good reference:

1816

Chinese diagnosis dolls

le cylindre

ECG - Atrial Fibrillation

Always

Increases with age and heart disease

Never

1. The most common significant arrhythmia
2. There are 23 professor emerita at Cal Poly Humboldt.
3. 1-3 of them have atrial fibrillation
4. Lose 15-30% of cardiac output
5. Increases risk for stroke
6. It is a hassle to get it identified.

The Heart Valves

TheHeart Sings

• Murmur is turbulence
• How valves can go wrong
• Complete failure

How did AI do it ?

I. Signal processing

1. Get rid of the noise

2. Segmentation - S₁ and S₂ and interval spaces in between

3. Spectral analysis - �

4. Mel-frequency - how the human ear hears noise.

Developed over evolution to a functional pattern.\

​

5. Cepstral coefficients (MFCCs).

The heart is an echo chamber

​

The terms "quefrency", "alanysis", "cepstrum" and "saphe" were invented by the authors by rearranging the letters in frequency, analysis, spectrum, and phase.

II. Deep neural networks

1. CNN - (convolutional neural network) - small filter - detects acoustic features

Human correlate - hardwired to identify edges, corners - pitch, pause

2. RNN - (recurrent neural network) use of memory - add in the element of time

in order to deal with dynamic phenomena

Human correlate - multiple memories

Sensory - iconic, echoic, haptic, smell, taste

Conscious - episodic, semantic (facts, language)

Unconscious - muscle memory, priming

3. LTSM -(long short term memory

Human correlate - short, working, long-term. Pruning through dreams.

4. Transformers act as attention focus

Human correlate - the brain as a Bayesian sibyl - generative models are updated by surprise

III. A clever plot twist - they used visual analysis of sonic frequencies

IV. Add ECG, ejection fraction, age, sex

Valvular Heart Disease

How did AI do ?

Overall Performance for Moderate-to-Severe Valvular Heart Disease

​

AI-enabled digital stethoscopes demonstrated 92.3% sensitivity compared to 46.2% - essentially doubling detection rates

​

However, specificity was slightly lower at 86.9% compared to 95.6% for clinicians News-Medical.

The technology represents a significant advance, particularly for primary care screening where

traditional auscultation frequently misses valvular disease in asymptomatic patients.

​

Clinical Implications

​

Best for severe disease - AI performs particularly well for severe AS and severe MR, where early detection has the greatest clinical impact

​

Screening tool, not replacement - These devices function best as an additional screening layer to support clinical judgment, not replace it

Cardiac output

1. Starling’s curve
2. Estimating ejection fraction
3. Estimating cardiac output. EF X HR. (Ejection fraction times heart rate)
4. Sensitivity 77-92%. Specificity 70-84%
5. Screening in primary care
6. Detecting asymptomatic heart failure
7. Rural / low-resource settings

Ejection fraction

Oh No !

_{More heart sounds}

S₁ and S_{2 are each two sounds}

S₃ and S₂

_Clicks

_{Artificial heart valves}

AI has to be able to distinguish artificial from normal

AI has to be able to distinguish abnormal from normal

• Older methods
• Ballistocardiography
• The tricorder

Ejection fraction

AI Programs Used for Heart Sound Analysis and Ejection Fraction

1. Gated Recurrent Unit (GRU): Gao et al. developed a heart failure screening framework using a GRU model that distinguished between normal subjects, heart failure with preserved ejection fraction (HFpEF),

and heart failure with reduced ejection fraction (HFrEF) using heart sounds, achieving 98.82% accuracy SciencePubMed Central.

2. CNN and RNN Combined

3. Deep Convolutional Generative Adversarial Networks (DCGAN): .

Two Networks in Competition

Two models are trained simultaneously by an adversarial process: a generator learns to create images that look real, while a discriminator learns to tell real images apart from fakes AJMC.

The concept is simple yet effective - the generator takes noise as input and generates data samples while the discriminator tries to classify those samples from real samples, causing a two-player game

where the generator tries to generate samples indistinguishable from real data while the discriminator becomes better at distinguishing between them

​

4 Wavelet Transform (DWT): Applied for denoising heart sound signals, as heart sounds are sensitive to noise and the DWT coefficient of heart sound signals is larger than noise components ScienceDirect.

Provides in formation about what frequencies and where they occur in time.

How did AI do it ?

How did AI do?

​

• 92.3% sensitivity for detecting moderate-to-severe valvular heart disease vs. 46.2% for traditional stethoscopes

​

• 98% accuracy for severe aortic stenosis and 94% for severe mitral regurgitation

Pulmonary Hypertension

• What is it - Stiffening of the blood vessels in the lungs - usually a low pressure system
• Why is it important ?
• What is the clinical significance of pulmonary hypertension ?
• How was it measured the old way ?
• How does an AI stethoscope estimate it ? (P₂ and timing)
• How good does the AI do ?

71% sensitivity means it would catch about 7 out of 10 cases of elevated pulmonary pressures, but miss roughly 3 out of 10

73% specificity means about 27% false positive rate - roughly 1 in 4 people without PH would screen positive