Abstract
This paper introduces "Flyweight" FLIM networks—extremely compact models (often under 100K parameters) designed for Salient Object Detection (SOD) in biomedical images. By using Feature Learning from Image Markers (FLIM) to learn kernels without backpropagation, the authors achieve competitive performance with heavyweight models while being orders of magnitude smaller and more efficient.
Introduction
Salient Object Detection (SOD) aims to identify the most visually distinct objects in an image. While deep learning has advanced SOD, it typically relies on massive annotated datasets and significant computational power—resources often unavailable in biomedical contexts like microscopy or MRI analysis. Furthermore, deploying large models on low-power medical devices is challenging.
Existing "lightweight" models often struggle with limited data because they still rely on backpropagation. This paper proposes a way to build highly specialized, efficient "flyweight" models that can be trained on just a handful of images by bypassing gradient-based weight updates in favor of marker-based feature discovery.
Methods
The core of this approach is FLIM (Feature Learning from Image Markers). Instead of using backpropagation, an expert provides "markers" (simple scribbles or dots) on a few representative images to identify foreground and background regions. The network extracts image patches from these regions and uses k-means clustering to define its convolutional kernels.
Several "Flyweight" innovations enhance this efficiency:
- Dilated-Separable Kernels: Standard convolutions are decomposed into depthwise and pointwise operations, with pointwise weights estimated from activation channel statistics rather than learned via backprop.
- Multi-Dilation Layers: Multiple dilation rates are used within the same layer to capture objects at different scales.
- Network Simplification: A novel pruning method uses a "uniqueness score" to remove redundant kernels, transferring their "energy" to remaining ones to maintain model magnitude.
- Adaptive Decoder: The encoder is paired with a decoder that estimates a pointwise convolution on-the-fly for every image, deciding which feature channels represent the object based on activation thresholds.
Results and Discussion
The researchers evaluated their models on two challenging datasets: S. mansoni parasite eggs (microscopy) and Brain Tumors (MRI). For both, models were trained using only five manually selected representative images.
Key results include:
- Extreme Efficiency: On the brain tumor dataset, the MDFLIM-S model was 14,000 times smaller* than U²Net in terms of parameter count (operating in the 1K–40K range vs. millions).
- Superior Performance: Despite its size, it achieved a weighted F-measure ($F_\beta^\omega$) of 0.739 on the tumor dataset, outperforming heavyweight models like U²Net (0.724) and BASNet.
- Speed: Training takes only a few seconds on a standard GPU and under 20 seconds on a CPU, compared to hours for backprop-based deep learning.
Conclusion
The paper demonstrates that for specialized biomedical tasks, clustering-based feature discovery (FLIM) can be as effective as gradient descent. By identifying redundancy and using domain-specific markers, the authors create "masters of one"—highly efficient models tailored for specific clinical or laboratory tools. Future work will address automating the selection of representative training images and increasing decoder robustness.
Summary updated