Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

A Deployment-Friendly Foundational Framework for Efficient Computational Pathology

Yu Cai, Cheng Jin, Jiabo Ma, Fengtao Zhou, Yingxue Xu, Zhengrui Guo, Yihui Wang, Zhengyu Zhang, Ling Liang, Yonghao Tan, Pingcheng Dong, Du Cai, On Ki Tang, Chenglong Zhao, Xi Wang, Can Yang, Yali Xu, Jing Cui, Zhenhui Li, Ronald Cheong Kin Chan, Yueping Liu, Feng Gao, Xiuming Zhang, Li Liang, Hao Chen, Kwang-Ting Cheng

TL;DR

LitePath introduces a deployment-friendly foundational framework for computational pathology that tightly couples a compact, distillation-based LiteFM with a lightweight Adaptive Patch Selector (APS) to mitigate over-parameterization and patch redundancy in gigapixel whole-slide images. By distilling knowledge from three state-of-the-art PFMs into LiteFM and employing APS for task-specific patch selection, LitePath achieves a 28x reduction in parameters and up to a 403.5x FLOP reduction relative to Virchow2, enabling edge-device deployment (e.g., Jetson Orin Nano Super at 25W) with strong accuracy. Across 37 cohorts and 26 tasks spanning four organs, LitePath attains competitive accuracy (average AUC retention of 99.71% vs Virchow2) and leads the Deployability Score (D-Score) compared to large PFMs, indicating superior balance between diagnostic performance and computational efficiency. Together, these results demonstrate that high-accuracy pathology AI can be rapidly deployed on accessible hardware, reducing energy consumption and enabling scalable, real-world clinical impact.

Abstract

Pathology foundation models (PFMs) have enabled robust generalization in computational pathology through large-scale datasets and expansive architectures, but their substantial computational cost, particularly for gigapixel whole slide images, limits clinical accessibility and scalability. Here, we present LitePath, a deployment-friendly foundational framework designed to mitigate model over-parameterization and patch level redundancy. LitePath integrates LiteFM, a compact model distilled from three large PFMs (Virchow2, H-Optimus-1 and UNI2) using 190 million patches, and the Adaptive Patch Selector (APS), a lightweight component for task-specific patch selection. The framework reduces model parameters by 28x and lowers FLOPs by 403.5x relative to Virchow2, enabling deployment on low-power edge hardware such as the NVIDIA Jetson Orin Nano Super. On this device, LitePath processes 208 slides per hour, 104.5x faster than Virchow2, and consumes 0.36 kWh per 3,000 slides, 171x lower than Virchow2 on an RTX3090 GPU. We validated accuracy using 37 cohorts across four organs and 26 tasks (26 internal, 9 external, and 2 prospective), comprising 15,672 slides from 9,808 patients disjoint from the pretraining data. LitePath ranks second among 19 evaluated models and outperforms larger models including H-Optimus-1, mSTAR, UNI2 and GPFM, while retaining 99.71% of the AUC of Virchow2 on average. To quantify the balance between accuracy and efficiency, we propose the Deployability Score (D-Score), defined as the weighted geometric mean of normalized AUC and normalized FLOP, where LitePath achieves the highest value, surpassing Virchow2 by 10.64%. These results demonstrate that LitePath enables rapid, cost-effective and energy-efficient pathology image analysis on accessible hardware while maintaining accuracy comparable to state-of-the-art PFMs and reducing the carbon footprint of AI deployment.

A Deployment-Friendly Foundational Framework for Efficient Computational Pathology

TL;DR

LitePath introduces a deployment-friendly foundational framework for computational pathology that tightly couples a compact, distillation-based LiteFM with a lightweight Adaptive Patch Selector (APS) to mitigate over-parameterization and patch redundancy in gigapixel whole-slide images. By distilling knowledge from three state-of-the-art PFMs into LiteFM and employing APS for task-specific patch selection, LitePath achieves a 28x reduction in parameters and up to a 403.5x FLOP reduction relative to Virchow2, enabling edge-device deployment (e.g., Jetson Orin Nano Super at 25W) with strong accuracy. Across 37 cohorts and 26 tasks spanning four organs, LitePath attains competitive accuracy (average AUC retention of 99.71% vs Virchow2) and leads the Deployability Score (D-Score) compared to large PFMs, indicating superior balance between diagnostic performance and computational efficiency. Together, these results demonstrate that high-accuracy pathology AI can be rapidly deployed on accessible hardware, reducing energy consumption and enabling scalable, real-world clinical impact.

Abstract

Pathology foundation models (PFMs) have enabled robust generalization in computational pathology through large-scale datasets and expansive architectures, but their substantial computational cost, particularly for gigapixel whole slide images, limits clinical accessibility and scalability. Here, we present LitePath, a deployment-friendly foundational framework designed to mitigate model over-parameterization and patch level redundancy. LitePath integrates LiteFM, a compact model distilled from three large PFMs (Virchow2, H-Optimus-1 and UNI2) using 190 million patches, and the Adaptive Patch Selector (APS), a lightweight component for task-specific patch selection. The framework reduces model parameters by 28x and lowers FLOPs by 403.5x relative to Virchow2, enabling deployment on low-power edge hardware such as the NVIDIA Jetson Orin Nano Super. On this device, LitePath processes 208 slides per hour, 104.5x faster than Virchow2, and consumes 0.36 kWh per 3,000 slides, 171x lower than Virchow2 on an RTX3090 GPU. We validated accuracy using 37 cohorts across four organs and 26 tasks (26 internal, 9 external, and 2 prospective), comprising 15,672 slides from 9,808 patients disjoint from the pretraining data. LitePath ranks second among 19 evaluated models and outperforms larger models including H-Optimus-1, mSTAR, UNI2 and GPFM, while retaining 99.71% of the AUC of Virchow2 on average. To quantify the balance between accuracy and efficiency, we propose the Deployability Score (D-Score), defined as the weighted geometric mean of normalized AUC and normalized FLOP, where LitePath achieves the highest value, surpassing Virchow2 by 10.64%. These results demonstrate that LitePath enables rapid, cost-effective and energy-efficient pathology image analysis on accessible hardware while maintaining accuracy comparable to state-of-the-art PFMs and reducing the carbon footprint of AI deployment.
Paper Structure (30 sections, 9 figures, 24 tables, 2 algorithms)

This paper contains 30 sections, 9 figures, 24 tables, 2 algorithms.

Table of Contents

  1. LitePath reduces theoretical overhead by two orders of magnitude compared to Virchow2.
  2. LitePath enables efficient deployment on resource-constrained edge devices.
  3. LitePath deployed on edge devices presents a scalable solution for high-speed, economical, and energy-efficient utilization of PFMs in clinical settings.
  4. LitePath demonstrates comparable performance to state-of-the-art PFMs across four cancer types.
  5. LitePath demonstrates strong potential for intraoperative analysis using frozen slides.
  6. LitePath demonstrates strong capability in pre-cancer diagnosis.
  7. WSIs contain substantial patch redundancy.
  8. APS can discard redundant patches while remaining non-inferior to the baseline in terms of AUC.
  9. The configuration for distillation pretraining of LiteFM is appropriate for both high accuracy and low computational overhead.
  10. Primary and Metastatic Cancer Classification.
  11. Cancer Subtyping.
  12. NSCLC Subtyping.
  13. Lymph Node Metastasis Prediction.
  14. IHC Status Prediction.
  15. TNM-N Staging (N0/N+).
  16. ...and 15 more sections

Figures (9)

  • Figure 1: Overview of the LitePath framework. LitePath is a deployment-friendly PFM framework designed to balance efficiency and diagnostic accuracy, consisting of LiteFM and APS. a, The inference pipeline of LitePath. LiteFM extracts features, while APS selects patches based on indices and shallow features $\{\mathbf{H}_i\}_{i=1}^N$ from block-1. Only the selected features $\{\mathbf{H}_s\}_{s \in \mathcal{S}}$ are propagated through the network for final prediction. b, LiteFM is distilled from Virchow2, H-Optimus-1 and UNI2 using approximately 190 million patches sourced from 72,280 WSIs. c, APS combines uniform sampling and attention-based sampling for patch selection. The scoring network is trained on the shallow features to approximate the attention score distribution of the final ABMIL. d, Average ranking scores based on Macro-AUC for 19 PFMs across all cohorts, internal cohorts, external cohorts, and prospective cohorts, respectively. e, Comparison of the prevailing deployment (Virchow2 on RTX 3090 GPUs) and the proposed solution (LitePath on Jetson Orin Nano Super devices) under an equivalent daily load and GPU budget.
  • Figure 2: Overall assessment of PFMs.a--d, Composition of the multi-center evaluation dataset for lung, breast, gastric, and colorectal cancers, respectively (H: Hospital). e, Trade-off between the average ranking score of Macro-AUC and model parameters. f, Trade-off between the average ranking score of Macro-AUC and throughput. g, Average D-Score of the PFMs. h, AUC retention of LitePath relative to Virchow2 across 37 cohorts. (#[No.]: task identifiers in the corresponding organ, with detailed mappings provided in panels a--d. Pros: prospective. Hospital identifiers are omitted for internal held-out test cohorts.)
  • Figure 3: Efficiency comparison of PFMs.a, Number of parameters for each PFM. LitePath (including LiteFM and APS modules) contains $28\times$ fewer parameters than Virchow2. b, Floating-point operations (FLOPs) required by each PFM to process a single input pathology patch. LitePath requires $38.8\times$ fewer FLOPs than Virchow2. c, Relative FLOPs of LitePath compared to LiteFM as a function of the number of patches per case. Only results for attention-based sampling are presented, as uniform-based sampling scales FLOPs in a straightforward linear fashion. For a sufficiently large number of patches, LitePath achieves a relative FLOPs convergence to 0.096, representing a $10.4\times$ reduction. Therefore, LitePath can theoretically deliver up to a 403.5-fold ($38.8\times10.4$) reduction in computational cost compared to Virchow2. d, Throughput of PFMs on RTX 3090 and Jetson Orin Nano Super GPUs, evaluated using dummy slides containing 30,000 patches with half-precision computation. "0.00" indicates out-of-memory. For these experiments, APS selected 1,000 patches using attention-based strategy, with no uniform selection performed.
  • Figure 4: Overall accuracy of PFMs across four organs.a--d, Performance of PFMs on lung, breast, gastric, and colorectal cancers, respectively. Each panel includes three visualizations: the average Macro-AUC across cohorts for the corresponding organ; the average ranking scores for all cohorts, internal cohorts, external cohorts (if applicable), and prospective cohorts (if applicable); and the distribution of ranking scores for each PFM.
  • Figure 5: Ablation study.a, Comparison of using all patches, using top-$k$ patches with highest final attention scores from ABMIL, and using uniform-$k$ patches. The average Macro-AUC across internal held-out test cohorts is presented. Detailed comparison on each task is provided in Extended Data Fig. \ref{['fig:topk']}. b, Non-inferiority test of the Adaptive Patch Selector (APS). The violin plots illustrate the distribution of AUC differences between LitePath (equipped with APS) and LiteFM (using all patches). For each plot, the mean value and 95% confidence interval (CI) of the AUC difference are displayed. The zero line and the non-inferiority margin (-2.5%) are indicated by gray and red dashed lines, respectively. Task identifier mappings are provided in Fig. \ref{['fig:results_overall']}a--d. c and d, Comparison of LiteFM family models based on average Macro-AUC and average ranking scores.
  • ...and 4 more figures