AI Research Portal

This paper introduces LowRank-CAM, a computationally efficient and interpretable framework for medical image analysis. It addresses the high computational cost of existing methods like Score-CAM by using Singular Value Decomposition (SVD) to extract dominant spatial modes from activation maps, thereby reducing the number of forward passes required for generating explanations. Experiments on the MURA dataset show that LowRank-CAM achieves a significant speedup over Score-CAM while maintaining comparable visual clarity and diagnostic relevance.

Existing Class Activation Map (CAM) methods, while providing visual explanations, can be computationally expensive, limiting their applicability in real-time or resource-constrained scenarios.
Perturbation-based CAM methods, such as Score-CAM, require a large number of forward passes through the classifier head for each activation map, leading to high computational cost. For K activation maps, Score-CAM requires (K + 1) forward passes with a complexity of O((K +1)Tf + KC).
This high computational demand hinders the integration of interpretable deep learning models into clinical decision-making processes where efficiency is crucial.

Proposes LowRank-CAM, a framework that aggregates activation maps into a global matrix and uses SVD to extract dominant spatial modes.
Reduces the number of forward passes required for generating explanations from K+1 to r (where r < K), significantly improving computational efficiency.
Demonstrates that LowRank-CAM achieves a 4.73 × speedup over Score-CAM on the MURA dataset while preserving visual clarity and diagnostic relevance.

LowRank-CAM constructs a global activation matrix from the final convolutional layer's activation maps. It then applies Singular Value Decomposition (SVD) to this matrix to extract dominant spatial patterns. These patterns, combined with class-specific scores, are used to generate cleaner and more efficient attention maps.

Workflow:

Extract Activations: Activation maps M ∈ R^(L×D×K) are extracted from the final convolutional layer.
Global Activation Matrix: Spatial dimensions are flattened, and all M are stacked to form Mg ∈ R^((L·D)×K).
SVD Decomposition: SVD is applied to Mg (Mg = UΣVᵀ) to retain the top r components (U, Σ, Vᵀ).
Low-rank Attention Mask Construction: For each retained rank-r, columns of U and rows of Vᵀ are used to form attention masks. These masks are then summed across channels after applying ReLU and normalization to produce clean low-rank attention masks. The formula for normalization is: A^(sp)(x, y) = Norm( Σ_k A_r(x, y, k) ).
Class Relevance Scoring: Each low-rank attention mask is applied to the activation maps via dot product, followed by global average pooling (GAP) and fully connected (FC) layers to compute class-specific weights (w_c). The formula used is: w_c = f_cls(M ⊙ A^(sp))[c].
Final CAM Generation: The final CAM is generated by combining the attention masks using the computed class-specific weights: L_LowRank-CAM(c) = Σ_r w_r A^(sp).

Key Advantage: LowRank-CAM reduces the computational cost from O((K+1)Tf) (Score-CAM) to O(KP² + rT_h), where P is the spatial dimension, r is the reduced rank, and T_h is the FC pass cost.

Architecture: The method is evaluated using the Inception-v3 architecture.
Key Equations:
- Normalization of attention masks: A^(sp)(x, y) = Norm( Σ_k A_r(x, y, k) )
- Class-specific weights computation: w_c = f_cls(M ⊙ A^(sp))[c]
- Final CAM generation: L_LowRank-CAM(c) = Σ_r w_r A^(sp)
Loss/Objective: Not stated in the paper.
Optimization: Not stated in the paper.
Augmentation: Not stated in the paper.
Data Handling: The MURA dataset of musculoskeletal radiographs was used.
Implementation Details: The rank (r) for LowRank-CAM was tuned using Insertion AUC and 1-Deletion AUC on ten wrist samples, varying ranks from 5 to 45. Rank 10 was chosen as optimal.

Datasets: MURA dataset (musculoskeletal radiographs), specifically the elbow study type.
Metrics: Average Drop (AD), Average Increase (AIn), Coherency (CH), Complexity (CX), and Average Drop-Complexity-Coherency (ADCC).
Baselines: CAM, Grad-CAM, Grad-CAM++, XGrad-CAM, Ablation-CAM, Eigen-CAM, Score-CAM, Score-CAM++, FIMF-Score-CAM, gScore-CAM, Group-CAM, Recipro-CAM.
Numeric Results:
- LowRank-CAM achieved an ADCC of 86.69 on the MURA elbow study using Inception-v3.
- This outperformed Score-CAM (82.07) and Grad-CAM (74.27).
- LowRank-CAM exhibited low complexity (CX: 24.30).
Runtime/Training Cost:
- LowRank-CAM achieved an inference time of 1.392 seconds per image.
- This provides a speedup of 4.73 × over Score-CAM (6.591 seconds).

No explicit paper-reported limitations were found.

Method evidence: "The resulting top-r low-rank attention masks, with r < K (r denotes the low-rank dimension and K is the total number of activation maps) replace per-channel perturbations and require only r forward passes through the classifier head." (Abstract)
Result evidence: "The MURA dataset of musculoskeletal radiographs (MURA dataset) was used to evaluate LowRank-CAM... As presented in Table 1, LowRank-CAM achieved an ADCC of 86.69, outperforming Score-CAM (82.07) and Grad-CAM (74.27). It also exhibited low complexity (CX: 24.30) and reduced inference time to 1.392 seconds, providing a speedup of 4.73 × over Score-CAM (6.591 seconds)." (Experiments and Results)
Limitation evidence: Not stated in the paper.

LowRank-CAM: A Computationally Efficient and Interpretable Framework for Medical Image Analysis (Student Abstract)

Summary

Problem

Key Contributions

Method

Technical Details

Results

Weaknesses / Limitations

Evidence

Additional Figures