📄 medicine

DNA-Inspired Symbolic Encoding with Adaptive Block-level Hybrid CGR-FCGR Features for Osteoporosis Progression Detection

This paper proposes a novel framework for osteoporosis screening using plain radiographs that combines entropy-guided adaptive block division, DNA-inspired symbolic encoding with CGR/FCGR features, and a hybrid HTGCA-Net deep learning model to achieve high accuracy in both binary and multiclass classification tasks.

Original authors: Lipika Dinda, Jitesh Pradhan, Manish Raj, Tapan Kumar Dey

Published 2026-07-03

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Lipika Dinda, Jitesh Pradhan, Manish Raj, Tapan Kumar Dey

Original paper licensed under CC BY 4.0 (https://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

The Big Picture: Finding Weak Bones Without the Expensive Machine

Imagine your bones are like a wooden fence. Over time, if the wood gets dry and brittle, the fence becomes weak and might break. This is osteoporosis.

Usually, doctors use a special, expensive machine called a DXA scanner to check how dense the wood is. But not every clinic has this machine, and it can be costly. This research team wanted to solve that problem. They asked: "Can we use the regular, cheap X-ray pictures that almost every hospital already has to spot weak bones?"

The answer they found is yes, but they had to build a very smart "digital detective" to do it.

The Problem: The Picture is Too Simple

Regular X-rays are just black-and-white photos. To a human eye (or a standard computer), the difference between healthy bone and weak bone is often very subtle—like trying to tell the difference between two shades of gray in a foggy photo. Standard computer programs often miss these tiny details.

The Solution: A Three-Step "DNA" Detective System

The researchers built a new system that treats the X-ray image not just as a picture, but as a secret code. Here is how their system works, step-by-step:

1. The "Smart Zoom" (Entropy-Guided Block Division)

Imagine you are looking at a map of a city. If you look at the whole city at once, you miss the details. If you look at every single house, you get overwhelmed.

What they did: Their system automatically scans the X-ray and asks, "Where is the interesting stuff?"
The Analogy: It's like a detective using a magnifying glass. If a part of the bone looks uniform and boring (like a smooth wall), the detective leaves it alone. But if a part looks messy, cracked, or complex (like a tangled knot of wood), the system zooms in and breaks that specific area into tiny, detailed pieces. This ensures they don't waste time on empty space and focus only on the "trouble spots."

2. Translating the Image into "DNA" (Symbolic Encoding)

Once they have zoomed in on the interesting parts, they need to read the pattern.

What they did: They took the gray pixels of the X-ray and turned them into a string of letters, like a sentence.
The Analogy: Imagine the X-ray is a secret message written in a code of 0s and 1s. The system translates those numbers into the four letters of DNA: A, C, G, and T.
- A dark pixel might become an "A".
- A light pixel might become a "T".
- Suddenly, the X-ray isn't a picture anymore; it's a long genetic story.

3. Reading the Story in Two Ways (CGR and FCGR)

Now that they have a "DNA story," they read it using two different methods to catch every clue.

Method A: The "Shape Map" (CGR)
- The Analogy: Imagine plotting the letters of your story on a square piece of paper. If you draw a line from the center to the corner representing "A," then to the corner for "C," and so on, you create a unique, fractal shape.
- Why it helps: Healthy bone creates a very organized, symmetrical shape. Weak, brittle bone creates a messy, scattered shape. This method captures the geometry of the bone.
Method B: The "Frequency Count" (FCGR)
- The Analogy: This is like counting how many times specific word combinations appear in a book. Does the word "A-T-G" appear a lot? Does "C-C-C" appear often?
- Why it helps: This captures the statistics. It tells the computer how often certain patterns repeat, which reveals the texture of the bone.

The Magic Mix: By combining the "Shape Map" and the "Frequency Count," the system gets a complete picture of the bone's health, seeing both the big picture and the tiny details.

The Brain: The "HTGCA-Net"

The researchers didn't just stop at reading the code; they built a super-smart brain to interpret it. They call this brain HTGCA-Net.

The Analogy: Think of this brain as a team of three experts working together:
1. The Local Observer (Convolution): Looks at small, immediate details.
2. The Big Picture Thinker (Transformer): Looks at how different parts of the image relate to each other over long distances.
3. The Context Keeper (GRU & Attention): Remembers the sequence of events and decides which clues are the most important to make a final decision.

This team works together to decide: Is this bone Normal, Osteopenic (weak but not broken yet), or Osteoporotic (very weak)?

The Results: How Good Was the Detective?

They tested this system on a large collection of real hip X-rays from Indian hospitals, where the "truth" was already known (checked by the expensive DXA machine).

Binary Test (Healthy vs. Broken): When asked to simply say "Healthy" or "Osteoporosis," the system was 98% accurate.
Three-Way Test (Healthy vs. Weak vs. Broken): When asked to distinguish between all three stages, it was 92% accurate.

The system was also very stable. Whether they gave it a lot of training data or a little, it kept performing well.

The Bottom Line

This paper claims that by turning X-ray images into "DNA-like" codes and using a special hybrid computer brain, they can detect osteoporosis with high accuracy using standard, cheap X-rays. This could eventually help doctors in places without expensive scanners to catch bone disease early.

Important Note: The authors state this is a screening tool. It helps flag potential problems, but it is not a replacement for the official medical diagnosis (the DXA scan) yet. It is designed to be a helpful, low-cost first step.

Technical Summary: DNA-Inspired Symbolic Encoding with Adaptive Block-level Hybrid CGR-FCGR Features for Osteoporosis Progression Detection

Problem Statement
Osteoporosis is a progressive skeletal disorder characterized by reduced bone mineral density (BMD) and deteriorating trabecular microarchitecture, leading to increased fracture risk. While Dual-Energy X-ray Absorptiometry (DXA) remains the clinical gold standard for diagnosis, its high cost, limited accessibility, and requirement for specialized infrastructure hinder widespread screening, particularly in low-resource settings. Although plain radiographs (X-rays) are widely available, traditional visual grading techniques suffer from subjectivity and reliance on clinical expertise. Existing deep learning approaches often lack interpretability, require massive annotated datasets, or focus on global representations that miss the minute local structural differences critical for early diagnosis. Furthermore, many current methods fail to effectively capture the complex, fractal-like alterations in trabecular bone patterns inherent to osteoporosis.

Methodology
The proposed framework introduces a novel pipeline that integrates entropy-guided spatial analysis, DNA-inspired symbolic encoding, and a hybrid deep learning architecture (HTGCA-Net) to classify osteoporosis from hip X-ray images. The methodology proceeds through the following stages:

Entropy-Guided Adaptive Block Division:
Instead of uniform grid partitioning, the system employs Shannon entropy to dynamically divide the preprocessed X-ray image into blocks. Regions exhibiting high textural complexity (high entropy) are recursively subdivided into finer $2 \times 2$ sub-blocks, while homogeneous regions are retained at a coarser level. This adaptive strategy focuses computational resources on diagnostically significant areas, such as trabecular thinning and cortical discontinuity, while reducing redundancy in uniform regions.
DNA-Inspired Symbolic Encoding:
To facilitate sequence-based analysis, grayscale pixel intensities are converted into symbolic sequences. Each pixel value (0–255) is represented as an 8-bit binary string, which is partitioned into four non-overlapping 2-bit segments. These segments are mapped to nucleotide symbols (A, T, G, C) using a predefined encoding rule (e.g., 00 $\to$ A, 01 $\to$ G, 10 $\to$ C, 11 $\to$ T). This transforms the image into a DNA-like symbolic sequence of length $L = 4 \times M \times N$ .
Hybrid Feature Extraction (CGR and FCGR):
The symbolic sequences are analyzed using two complementary techniques:
- Chaos Game Representation (CGR): This maps the symbolic sequence into a 2D coordinate space to capture spatial recurrence and fractal patterns. The resulting coordinate distribution is discretized into a density matrix, capturing spatial structural characteristics and local intensity transitions.
- Frequency Chaos Game Representation (FCGR): This quantifies the frequency distribution of $k$ -mer subsequences (contiguous nucleotide combinations) within the symbolic sequence. The FCGR matrix provides a statistical, frequency-domain perspective of the bone texture, capturing global dependencies and repetitive structural motifs.
- Feature Fusion: The spatial descriptors from CGR and the frequency-based statistics from FCGR are concatenated to form a high-dimensional hybrid feature vector (71,680 features), integrating both local structural variations and global statistical patterns.
HTGCA-Net Architecture:
The extracted hybrid features are processed by a Hybrid Transformer-GRU Context Attention Network (HTGCA-Net). The architecture consists of:
- Conv1D Layer: Extracts local dependencies among neighboring features.
- Transformer Encoder: Utilizes multi-head self-attention to model long-range global dependencies within the feature sequence.
- Bidirectional GRU (BiGRU): Captures sequential context from both forward and backward directions.
- Context-Attention Mechanism: Assigns importance weights to features based on a trainable context vector, focusing the model on the most discriminative regions.
- Classification Head: A fully connected layer with Softmax (for multiclass) or Sigmoid (for binary) activation for final prediction.

Key Contributions
The paper outlines five primary contributions:

Entropy-Guided Adaptive Partitioning: A dynamic block division strategy that refines regions of high textural irregularity, enhancing the representation of local trabecular variations.
DNA-Inspired Symbolic Encoding: A method to convert grayscale intensity patterns into nucleotide sequences, enabling the application of fractal and sequence-based analysis techniques to medical images.
Multi-Domain Feature Representation: The integration of CGR (spatial/fractal) and FCGR (frequency/statistical) to provide a comprehensive description of trabecular bone morphology.
Fusion Strategy: A hybrid feature fusion approach that combines spatial recurrence patterns with frequency-based symbolic statistics to improve discriminative capability between normal, osteopenic, and osteoporotic states.
HTGCA-Net: A novel deep learning architecture that synergizes convolutional layers, Transformer encoders, BiGRU units, and context attention to learn hierarchical dependencies for robust binary and multiclass classification.

Experimental Results
The framework was evaluated on a clinically validated dataset of 1,678 hip radiographs collected under an ICMR project, with labels derived from WHO T-score criteria (Normal, Osteopenia, Osteoporosis). Experiments were conducted using patient-wise splits (60:40, 70:30, 80:20, 90:10) to ensure no data leakage.

Binary Classification (Normal vs. Osteoporosis): Using an 80:20 split, the hybrid feature approach achieved an accuracy of 0.98, with a macro-averaged F1-score of 0.99. This outperformed standalone CGR and FCGR features, which achieved accuracies of approximately 0.97.
Multiclass Classification (Normal, Osteopenia, Osteoporosis): The hybrid approach achieved an accuracy of 0.92 under the 80:20 split, with macro-averaged F1-scores of 0.92.
Ablation Study: Removing key components (Transformer, BiGRU, or Attention) resulted in performance drops of 2–4%, confirming the necessity of each module. The baseline model (Conv1D + Dense) showed a significant degradation (approx. 15% drop in binary accuracy), highlighting the importance of the hybrid architecture.
Comparison: The proposed method demonstrated competitive or superior performance compared to existing state-of-the-art methods (e.g., ResNet, EfficientNet, and various ensemble CNNs) in terms of sensitivity, specificity, and accuracy.

Significance and Claims
The paper claims that the proposed framework offers a reliable, cost-effective, and scalable solution for osteoporosis screening using widely available plain radiographs. By combining entropy-guided spatial analysis with DNA-inspired symbolic encoding and hybrid deep learning, the system effectively captures subtle morphological differences that are often missed by conventional global or pixel-level approaches.

The authors emphasize that the hybrid CGR-FCGR feature representation provides a more robust and discriminative description of bone texture than individual feature sets, leading to stable performance across different data distributions. The inclusion of attention mechanisms enhances interpretability by highlighting clinically significant structural changes. The study positions the system as a supportive screening tool intended to improve accessibility to osteoporosis detection, particularly in settings where DXA is unavailable, rather than as a direct replacement for quantitative BMD measurement.

Limitations and Future Work
The authors acknowledge that the current framework focuses on discrete class prediction rather than direct BMD estimation. Additionally, the symbolic encoding may not capture all complex anatomical variations, and the model's complexity requires optimization for resource-constrained environments. Future work aims to extend the framework to regression-based modeling for direct BMD estimation, incorporate clinically interpretable visualization techniques, and validate the method on larger, multi-center datasets with diverse demographics.