Abstract
To improve the efficacy of traditional medical image retrieval, this paper presents a new medical algorithm that uses discrete cosine transformation (DCT) innovatively. The study utilized a database of 42 upper torso CT images from patients with hepatocellular carcinoma at ShengJing Hospital. By comparing retrieval results—specifically the similarity of six retrieved images to a query image—the efficacy of different algorithms was evaluated. Results indicate that DCT algorithms display significantly better retrieval performance than traditional grayscale histogram cumulative algorithms, extracting more visually and mathematically similar images. While some drawbacks remain, this technology shows promising prospects for integration into hospital medical systems.
The Challenge of Spatial Blindness in CBIR
Content-Based Image Retrieval (CBIR) is a cornerstone of modern medical diagnostics, allowing clinicians to find historical cases similar to a current patient's scans. However, traditional methods often rely on grayscale histograms, which merely tally pixel intensities. This approach is "spatially blind"; two images with identical pixel counts but completely different anatomical structures (e.g., a tumor in the left lobe vs. the right) would appear identical to the algorithm. In clinical settings, where the spatial arrangement of tissues is critical, this leads to poor accuracy and irrelevant matches.
Methodology: A Two-Stage Hybrid Approach
To address these limitations, this research proposes a hybrid system combining segmented histograms with the Discrete Cosine Transform (DCT).
1. Segmented Histogram Pre-filtering
To introduce basic spatial awareness, the image is divided into four equal quadrants. Histograms are calculated for each quadrant individually, and the results are averaged. This ensures that a match must share similar pixel distributions in the same general regions of the image. This stage acts as a "pre-filter" with a similarity threshold of 0.75 to discard obviously irrelevant data.
2. Feature Extraction via DCT
The core innovation lies in applying DCT to the remaining candidates. Unlike the Discrete Fourier Transform (DFT), DCT uses only real numbers, reducing computational complexity. More importantly, DCT excels at "energy concentration," pushing the most significant structural information into low-frequency coefficients. By comparing these coefficients using Euclidean distance, the system identifies images that share structural "essence" rather than just brightness profiles.
Experimental Results
The algorithm was tested on a dataset of 42 DICOM-format CT images. The DCT-based approach demonstrated a marked improvement in retrieving images that align with human visual criteria.
Quantitative analysis confirms the superiority of the DCT method across three key metrics: specificity, accuracy, and sensitivity.
As shown in the data, the DCT algorithm achieved a specificity of 94.6%, nearly 34 percentage points higher than the traditional histogram method. Accuracy also saw a significant jump from 38.4% to 66.7%.
Discussion and Future Work
While the DCT algorithm significantly outperforms traditional methods, it is not without limitations. The primary hurdle is computational speed; the process is currently too slow for real-time clinical applications. Furthermore, the small dataset (42 images) necessitates further testing on larger, more diverse databases to ensure scalability. Future research will focus on optimizing the mathematical operations to achieve rapid response times and developing more objective tools for evaluating image similarity.
Despite these challenges, the shift from simple pixel counting to energy-based structural analysis represents a major step forward in medical image retrieval technology.
Summary updated