🤖 AI

PRISM: Perinuclear Ring-based Image Segmentation Method for Acute Lymphoblastic Leukemia Classification

The paper introduces PRISM, a novel method for classifying Acute Lymphoblastic Leukemia that bypasses the challenges of cytoplasmic segmentation by using adaptive perinuclear rings to extract robust texture and color descriptors, achieving high accuracy (98.46%) and precision-recall AUC (0.9937) through a calibrated ensemble of traditional classifiers.

Original authors: Larissa Ferreira Rodrigues Moreira, Leonardo Gabriel Ferreira Rodrigues, Rodrigo Moreira, André Ricardo Backes

Published 2026-05-14

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Larissa Ferreira Rodrigues Moreira, Leonardo Gabriel Ferreira Rodrigues, Rodrigo Moreira, André Ricardo Backes

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 are a doctor trying to spot a specific type of troublemaker (a cancerous blood cell) in a crowded room full of innocent people (healthy blood cells). The troublemakers look very similar to the innocent ones, and the lighting in the room is often dim or uneven, making it hard to tell them apart.

For a long time, computers trying to do this job had two main problems:

The "Outline" Problem: They tried to draw a perfect circle around every single cell to separate it from the background. But because the cells are sticky, overlapping, and the colors are messy, the computer kept getting confused and drawing the wrong lines.
The "Black Box" Problem: Newer, smarter computers (Deep Learning) stopped trying to draw lines and just guessed based on the whole picture. They got good at guessing, but they were like a magic 8-ball: they gave the right answer, but no one knew why they said yes or no. Plus, they needed massive, expensive supercomputers to learn.

Enter PRISM: The "Concentric Ring" Detective

The authors of this paper, Larissa, Leonardo, Rodrigo, and André, built a new method called PRISM. Instead of trying to draw the messy outer edge of the cell, they decided to focus on the one thing that is always stable and easy to find: the nucleus (the dark center of the cell).

Here is how PRISM works, using a simple analogy:

1. The Anchor (The Nucleus)

Think of the nucleus as a lighthouse in a foggy sea. It's the only part of the cell that is dark, round, and easy to spot. PRISM finds this lighthouse first and locks onto it.

2. The Rings (The Perinuclear Zones)

Instead of trying to find the shoreline (the cell membrane), PRISM draws two invisible, expanding rings around the lighthouse:

Ring 1 (The Proximal Zone): This is the immediate area right next to the lighthouse. In healthy cells, this area looks one way; in sick cells, it looks different.
Ring 2 (The Distal Zone): This is a slightly wider ring further out.

By looking at these specific rings, the computer doesn't need to know where the cell ends and the background begins. It just looks at the "neighborhood" right around the nucleus.

3. The Clues (Features)

Once the rings are drawn, PRISM acts like a forensic scientist. It doesn't just look at the picture; it measures specific clues:

Colors: How blue or yellow is the ring? (Cancer cells often have different color patterns).
Texture: Is the area smooth or grainy? (Like comparing smooth silk to rough sandpaper).
Gradients: How quickly does the color change as you move from the center out to the ring?

It turns all these visual clues into a simple list of numbers (a "table"), rather than trying to process a giant, complex image.

4. The Panel of Judges (The Ensemble)

Finally, PRISM doesn't rely on just one computer program to make the final decision. It uses a "team" of different, simpler algorithms (like a team of detectives with different specialties).

Some are good at spotting patterns in trees (Random Forest).
Some are good at drawing lines between groups (SVM).
They all vote, and a "head judge" (a meta-classifier) looks at their votes to make the final call.

The Results: Why is this a big deal?

The paper tested this method on a standard dataset of blood cells (ALL-IDB2) and found some impressive things:

It's incredibly accurate: PRISM got 98.46% accuracy. This means it correctly identified the sick cells almost every time.
It beats the "Black Boxes": Surprisingly, this simple, ring-based method performed better than or equal to the most advanced, complex Deep Learning models (like ResNet and Swin Transformers) that require massive supercomputers.
It's transparent: Because PRISM looks at specific rings and colors, a doctor can actually understand why the computer made a decision. It's not a magic guess; it's based on measurable facts.
It's cheap and fast: You don't need a $10,000 graphics card to run this. It runs easily on a standard computer (CPU only), making it perfect for places with limited resources.

In Summary
PRISM is a clever workaround. Instead of fighting a losing battle to draw perfect outlines around messy cells, it ignores the edges entirely. It focuses on the stable center, measures the immediate surroundings, and uses a team of simple, explainable tools to diagnose the problem. It proves that sometimes, you don't need a super-complex AI to solve a medical problem; you just need a smart, focused approach.

Technical Summary: PRISM for Acute Lymphoblastic Leukemia Classification

Problem Statement
The automated diagnosis of Acute Lymphoblastic Leukemia (ALL) via peripheral blood smear analysis faces significant hurdles. Traditional Computer-Aided Diagnosis (CAD) pipelines rely heavily on precise cell membrane segmentation to extract morphological biomarkers (e.g., nuclear enlargement, cytoplasmic reduction). However, this step frequently fails in routine laboratory conditions due to overlapping erythrocytes, low nucleocytoplasmic contrast, and staining inconsistencies. Conversely, while Deep Learning (DL) approaches (CNNs, Vision Transformers) have achieved high accuracy, they often function as "black boxes" lacking interpretability, are prone to overfitting on small medical datasets, and require substantial computational resources (GPUs) that limit deployment in resource-constrained or decentralized healthcare settings.

Methodology: The PRISM Framework
The authors propose the Perinuclear Ring-based Image Segmentation Method (PRISM), a lightweight, GPU-independent approach that circumvents the need for explicit full-cell membrane delineation. The methodology is structured into four integrated phases:

Preprocessing and Nuclear Segmentation:
- Input images (256 × 256) are standardized and mapped to CIELAB and HSV color spaces.
- A custom chromatic scoring map ( $S_{chroma}$ ) is generated to maximize nucleocytoplasmic contrast by weighting the blue-yellow opponent channel and saturation.
- The nucleus is segmented using Otsu's thresholding on the lightness channel, with a K-Means clustering fallback for edge cases. Topological heuristics filter artifacts, resulting in a robust nuclear mask ( $M_n$ ) that serves as the geometric anchor.
PRISM-Based Spatial Mapping (Zonal Extraction):
- Instead of segmenting the entire cell, PRISM constructs adaptive, concentric zones around the nucleus ( $M_n$ ).
- An adaptive dilation strategy calculates an expansion radius ( $\delta^*$ ) based on the nuclear area to dynamically bound the zones, preventing overextension.
- Two mutually exclusive spatial zones are defined:
  - $Z_1$ (Proximal Interface): The immediate perinuclear region ( $D_{\delta^*_1}(M_n) \setminus M_n$ ).
  - $Z_2$ (Distal Cytoplasm): The region further from the nucleus ( $D_{\delta^*_2}(M_n) \setminus D_{\delta^*_1}(M_n)$ ).
- These zones, along with the nucleus and the gross cell boundary, serve as independent bases for feature extraction.
Multidomain Feature Extraction:
- Morphological/Topological: Geometric properties (area, circularity, solidity) and the Nucleocytoplasmic (N:C) ratio.
- Chromatic Statistics: First-order moments (mean, standard deviation) of RGB and grayscale channels. Crucially, the method computes spatial intensity gradients ( $\Delta$ ) representing the decay of color and texture from the nucleus to the periphery ( $\mu(M_n) - \mu(Z_i)$ ).
- Textural Analysis: Macro-texture features (Contrast, Homogeneity, Energy, Correlation) derived from Gray Level Co-occurrence Matrices (GLCM) and micro-texture heterogeneities captured via Local Binary Patterns (LBP), downsampled to 32 bins to reduce dimensionality.
Meta-Classification (Ensemble Stacking):
- A heterogeneous ensemble stacking architecture is employed.
- Level-0: A diverse pool of base classifiers (Random Forest, Extra Trees, XGBoost, SVM, KNN, Logistic Regression).
- Calibration: Platt Scaling is applied to base model outputs to ensure they represent true posterior probabilities.
- Level-1: A Support Vector Machine (SVM) with an RBF kernel acts as the meta-classifier, trained on Out-Of-Fold (OOF) probabilities to prevent data leakage.

Key Contributions

Adaptive Mathematical Methodology: PRISM replaces the error-prone task of full-cell segmentation with a multiscale, concentric zonal approach anchored on the biologically stable nucleus.
Interpretability: Unlike opaque DL models, PRISM grounds predictions in explicit, clinically verifiable morphological and chromatic gradients at the nucleocytoplasmic interface.
Computational Efficiency: The framework translates a high-dimensional computer vision task into a lightweight tabular machine learning problem, enabling high-performance inference on CPU-only, resource-constrained devices without GPU dependency.
State-of-the-Art Performance: Experimental validation demonstrates that the PRISM-driven ensemble outperforms or matches state-of-the-art end-to-end DL architectures on the ALL-IDB2 dataset.

Experimental Results
Evaluated on the ALL-IDB2 dataset (260 isolated cells: 130 leukemic, 130 healthy) using 5-fold stratified cross-validation:

Performance: The optimal ensemble configuration (Extra Trees + SVM + Logistic Regression) achieved an accuracy of 98.46%, a Precision-Recall AUC of 0.9937, and a Matthews Correlation Coefficient (MCC) of 0.9698.
Ablation Study: The study found that a minimal ensemble of two models (RF + SVM) was sufficient to reach peak accuracy (98.46%), while increasing complexity beyond three models offered no significant gains and occasionally degraded performance due to meta-classifier overfitting.
Comparison with DL: PRISM outperformed five DL baselines (ResNet-18, VGG16, DenseNet-121, EfficientNet-B0, Swin Transformer). Notably, EfficientNet-B0 achieved 98.08% accuracy, while PRISM achieved 98.46% without requiring pre-trained weights, extensive transfer learning, or GPU acceleration.
Comparison with Literature: PRISM matches the performance of recent complex DL ensembles (e.g., [Rodrigues et al. 2022], [Das et al. 2025]) and significantly outperforms traditional handcrafted methods, all while maintaining a low-cost, GPU-less execution pipeline.

Significance and Claims
The paper claims that PRISM establishes a new standard for automated lymphoblast screening by balancing accuracy, interpretability, and efficiency. Its primary significance lies in:

Biological Focus: By geometrically restricting analysis to the nucleus and perinuclear rings, the method avoids learning confounding background artifacts (e.g., staining variations, erythrocyte density) that often plague end-to-end CNNs.
Clinical Viability: The framework provides a transparent decision-making process based on verifiable biological markers, addressing the "black box" limitation of DL in medical environments.
Accessibility: By eliminating the need for high-end GPUs and complex training pipelines, PRISM offers a scalable solution suitable for point-of-care diagnostics in resource-limited or decentralized settings.

The authors conclude that segmentation-light, interpretable representations can deliver high-accuracy ALL screening with practical efficiency, and they propose future work to integrate PRISM into portable microscopy pipelines and validate it on real-world clinical data from diverse laboratories.