🤖 AI

Quantized Machine Learning Models for Medical Imaging in Low-Resource Healthcare Settings

This paper presents a multi-strategy compression framework for brain tumor classification that demonstrates, through full experimental validation of a Float16 quantized MobileNetV2 model, the ability to achieve a 6.14x reduction in model size with negligible accuracy loss, thereby enabling clinically viable deployment in low-resource healthcare settings.

Original authors: Sumanth Meenan Kanneti, Aryan Shah

Published 2026-05-20

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Sumanth Meenan Kanneti, Aryan Shah

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a brilliant, highly educated doctor who can look at a brain scan and instantly tell you if there is a tumor, what kind it is, or if the brain is healthy. This "doctor" is actually a computer program called a Deep Learning Model.

However, this doctor is incredibly heavy. Not physically, but in terms of computer memory. He requires a massive, expensive supercomputer to live in. If you try to bring this doctor to a small clinic in a remote village that only has a basic laptop or a tablet, he won’t fit. The computer will crash, or the battery will die instantly.

This paper is about shrinking that brilliant doctor so he can fit into a small backpack, without losing his medical knowledge.

The Problem: The "Heavy" Doctor

The authors focused on a specific task: looking at MRI brain scans to classify four things:

Glioma (a type of brain tumor)
Meningioma (another type of tumor)
Pituitary tumor (a tumor near the base of the brain)
Healthy brain (no tumor)

The original computer model used to do this was accurate, but it was huge—taking up 35 megabytes of space. That’s like trying to carry a full library’s worth of books when you only need a single novel. It’s too big for low-power devices like Raspberry Pis or mobile phones often used in resource-poor areas.

The Solution: "Quantization" (The Art of Summarizing)

The authors used a technique called Quantization.

Think of it like this: The original model writes its notes in extremely precise, long-winded sentences (called Float32). It says, "The pixel intensity is exactly 0.123456789..."

Quantization teaches the model to write in shorthand (called Float16). It says, "The pixel intensity is about 0.12..."

By rounding off the numbers, the model becomes much smaller and faster to read. The fear, of course, is that by rounding off the numbers, the doctor might make mistakes. Will he miss a tumor because he wasn’t precise enough?

The Experiment: Testing the Shorthand Doctor

The team built a lightweight version of the model using an architecture called MobileNetV2 (think of it as a streamlined, efficient car engine rather than a heavy truck engine). They trained it to recognize the four brain conditions.

Then, they applied the "shorthand" (Float16 quantization) to shrink it.

The Results:

Size: The model shrank from 35.34 MB down to 5.76 MB. That is a 6.14 times reduction. It’s like compressing a suitcase so it fits in a handbag.
Accuracy: This is the most important part. The original "precise" doctor got 82.20% of the diagnoses right. The new "shorthand" doctor got 82.37% right.

Surprisingly, the shorthand doctor was slightly more accurate (though the difference is so small it’s basically the same). This means the model didn’t lose any medical knowledge by using shorthand.

Did the Shorthand Doctor Miss Anything Specific?

The authors checked if the shorthand caused the doctor to be bad at recognizing specific types of tumors.

Healthy Brains: He was very good at spotting these (89% success rate).
Pituitary Tumors: He was also very good (86% success rate).
Gliomas: He was quite good (85% success rate).
Meningiomas: This was the hardest one (72% success rate), but interestingly, the shorthand didn’t make it worse. The model struggled with Meningiomas in the same way before and after shrinking.

The main confusion was between Meningiomas and Gliomas, which look very similar in MRI scans. But the compression didn’t create new mistakes; it just kept the existing ones.

What About Other Methods?

The paper mentions two other ways to shrink the model:

Quantization-Aware Training (QAT): Teaching the model to write in shorthand while it is still learning, rather than converting it afterward.
Knowledge Distillation: Having a huge, expert teacher model (DenseNet-101) teach a small student model (DenseNet-32) what it knows.

However, the authors did not test these two methods in this specific paper. They only fully tested the "MobileNetV2 with Float16 shorthand" method. They say they will test the other two methods in future work.

The Bottom Line

The paper proves that you can take a complex, heavy AI brain scanner and compress it into a tiny, lightweight version that fits on cheap, low-power hardware. It does this without losing any diagnostic ability.

In simple terms: They turned a heavy, precise encyclopedia into a concise, accurate pocket guide. The pocket guide takes up 6 times less space, fits in your pocket (or a small clinic’s computer), and gives you the same correct answers as the encyclopedia. This makes it possible to use advanced AI for brain tumor screening in places that don’t have expensive supercomputers.

Technical Summary: Quantized Machine Learning Models for Medical Imaging in Low-Resource Healthcare Settings

Problem Statement
Deep learning models, particularly Convolutional Neural Networks (CNNs), have demonstrated high performance in medical image analysis, including brain tumor classification from MRI scans. However, deploying these models in low-resource clinical environments—such as rural clinics, mobile diagnostic units, and under-resourced hospitals—is hindered by computational, memory, and power constraints. Most high-performing models require GPUs, large memory footprints, and stable power supplies, creating a gap between algorithmic capability and deployable clinical utility. There is a need for efficient models that preserve diagnostic performance while operating under strict compute and memory constraints suitable for edge devices.

Methodology
The authors propose a multi-strategy compression framework for multi-class brain tumor classification (glioma, meningioma, pituitary tumor, and healthy controls) from MRI images. The framework encompasses three complementary approaches:

Float16 Post-Training Quantization (PTQ): Applied to a lightweight MobileNetV2 backbone. This is the primary pipeline experimentally validated in this work.
Quantization-Aware Training (QAT): A strategy using the TensorFlow Model Optimization Toolkit to simulate low-precision arithmetic during training for INT8 conversion.
Knowledge Distillation: A pipeline where a compact DenseNet-32 student model learns from a high-capacity DenseNet-101 teacher model, followed by low-bit PTQ.

Note: The paper explicitly states that the complete empirical evaluation of the QAT and DenseNet distillation strategies is reserved for future work. The following details focus on the experimentally validated MobileNetV2 pipeline.

Data and Preprocessing
The study utilizes a publicly available dataset containing 2,870 training and 609 test images of contrast-enhanced T1-weighted MRI slices. The data is split into training, validation (20% stratified split from training data), and test sets. Images are preprocessed by converting grayscale to RGB, resizing to 224×224 pixels, and normalizing pixel intensities to the [0,1] range. Data augmentation (random rotation, flipping, shifting, zooming, and brightness perturbation) is applied only to training images to improve generalization.

Model Architecture and Training
The primary model uses a MobileNetV2 backbone pretrained on ImageNet, chosen for its efficiency in edge applications. A custom classification head is attached, consisting of global average pooling, batch normalization, two dense layers (512 and 256 units) with ReLU activation, L2 regularization, and dropout, followed by a softmax output layer for four classes.

The model is trained via a three-stage transfer learning process:

Stage 1: The MobileNetV2 backbone is frozen; only the classification head is trained for up to 15 epochs with a learning rate of $10^{-3}$ .
Stage 2: The entire network is unfrozen and fine-tuned end-to-end for up to 30 epochs with a reduced learning rate of $10^{-5}$ .
Stage 3: Training continues from the best Stage 2 checkpoint for up to 20 additional epochs with an ultra-low learning rate of $5 \times 10^{-6}$ .

Results
The full-precision Float32 baseline model achieved a validation accuracy of 82.20%. After applying Float16 post-training quantization via TensorFlow Lite, the quantized model achieved a validation accuracy of 82.37%, representing a negligible difference ( $-0.17\%$ in favor of the quantized model) that falls within normal evaluation variance.

Compression Efficiency: The model size was reduced from 35.34 MB (Float32) to 5.76 MB (Float16), achieving a compression ratio of 6.14×. This exceeds the typical 2× reduction expected from bit-width halving alone, due to graph optimizations and the removal of training-specific nodes during TFLite conversion.
Per-Class Performance: Quantization preserved diagnostic performance uniformly across all four tumor categories. The model performed best on the "No Tumor" class (F1 = 0.89) and "Pituitary" class (F1 = 0.86). "Meningioma" was the most challenging class (F1 = 0.73).
Error Analysis: The confusion matrix revealed that the most frequent misclassifications occurred between meningioma and glioma (28 cases) and between pituitary and meningioma (25 cases), consistent with known radiological overlaps in T1-weighted MRI. These error patterns remained virtually identical before and after quantization.

Significance and Contributions
The paper claims three main contributions:

Framework Design: It presents a multi-strategy compression framework integrating PTQ, QAT, and knowledge distillation for brain tumor MRI classification, detailing the methodology for each.
Experimental Validation: It demonstrates that Float16 PTQ on a MobileNetV2 backbone achieves a 6.14× reduction in model size with no meaningful accuracy loss, producing a 5.76 MB model suitable for CPU-only and low-power edge devices.
Clinical Viability: It confirms through per-class evaluation that quantization preserves diagnostic performance uniformly across all tumor categories, supporting the feasibility of lightweight quantized models for reliable AI-assisted diagnostics in resource-constrained healthcare settings.

The authors emphasize that this work contributes practical evidence toward democratizing medical imaging AI by enabling deployment on hardware commonly found in low-resource environments, such as ARM-based microcontrollers and single-board computers, without requiring GPU hardware.