A Model-Agnostic Approach to Mitigate Gradient Interference for Multi-Task Learning

pFedBCC: Personalizing Federated multi-target domain adaptive segmentation via Bi-pole Collaborative Calibration

BACKGROUND AND OBJECTIVE

Multi-target domain adaptation (MTDA) is a well-established technology for unsupervised segmentation. It can significantly reduce the workload of large-scale data annotations, but assumes that each domain data can be freely accessed. However, data privacy limit its deployment in real-world medical scenes. Aiming at this problem, federated learning (FL) commits a paradigm to handle private cross-institution data.

METHODS

This paper makes the first attempt to apply FedMTDA to medical image segmentation by proposing a personalized Federated Bi-pole Collaborative Calibration (pFedBCC) framework, which leverages unannotated private client data and a public source-domain model to learn a global model at the central server for unsupervised multi-type immunohistochemically (IHC) image segmentation. Concretely, pFedBCC tackles two significant challenges in FedMTDA including client-side prediction drift and server-side aggregation drift via Semantic-affinity-driven Personalized Label Calibration (SPLC) and Source-knowledge-oriented Consistent Gradient Calibration (SCGC). To alleviate local prediction drift, SPLC personalizes a cross-domain graph reasoning module for each client, which achieves semantic-affinity alignment between high-level source- and target-domain features to produce pseudo labels that are semantically consistent with source-domain labels to guide client training. To further alleviate global aggregation drift, SCGC develops a new conflict-gradient clipping scheme, which takes the source-domain gradient as a guidance to ensure that all clients update with similar gradient directions and magnitudes, thereby improving the generalization of the global model.

RESULTS

pFedBCC is evaluated on private and public IHC benchmarks, including the proposed MT-IHC dataset, and the panCK, BCData, DLBC-Morph and LYON19 datasets. Overall, pFedBCC achieves the best performance of 88.8% PA on MT-IHC, as well as 88.4% PA on the LYON19 dataset, respectively.

CONCLUSIONS

The proposed pFedBCC performs better than all comparison methods. The ablation study also confirms the contribution of SPLC and SCGC for unsupervised multi-type IHC image segmentation. This paper constructs a MT-IHC dataset containing more than 19,000 IHC images of 10 types (CgA, CK, Syn, CD, Ki67, P40, P53, EMA, TdT and BCL). Extensive experiments on the MT-IHC and public IHC datasets confirm that pFedBCC outperforms existing FL and DA methods.

Multi-Instance Multi-Task Learning for Joint Clinical Outcome and Genomic Profile Predictions From the Histopathological Images

With the remarkable success of digital histopathology and the deep learning technology, many whole-slide pathological images (WSIs) based deep learning models are designed to help pathologists diagnose human cancers. Recently, rather than predicting categorical variables as in cancer diagnosis, several deep learning studies are also proposed to estimate the continuous variables such as the patients' survival or their transcriptional profile. However, most of the existing studies focus on conducting these predicting tasks separately, which overlooks the useful intrinsic correlation among them that can boost the prediction performance of each individual task. In addition, it is sill challenge to design the WSI-based deep learning models, since a WSI is with huge size but annotated with coarse label. In this study, we propose a general multi-instance multi-task learning framework (HistMIMT) for multi-purpose prediction from WSIs. Specifically, we firstly propose a novel multi-instance learning module (TMICS) considering both common and specific task information across different tasks to generate bag representation for each individual task. Then, a soft-mask based fusion module with channel attention (SFCA) is developed to leverage useful information from the related tasks to help improve the prediction performance on target task. We evaluate our method on three cancer cohorts derived from the Cancer Genome Atlas (TCGA). For each cohort, our multi-purpose prediction tasks range from cancer diagnosis, survival prediction and estimating the transcriptional profile of gene TP53. The experimental results demonstrated that HistMIMT can yield better outcome on all clinical prediction tasks than its competitors.

Turing patterns in simplicial complexes

The spontaneous emergence of patterns in nature, such as stripes and spots, can be mathematically explained by reaction-diffusion systems. These patterns are often referred as Turing patterns to honor the seminal work of Alan Turing in the early 1950s. With the coming of age of network science, and with its related departure from diffusive nearest-neighbor interactions to long-range links between nodes, additional layers of complexity behind pattern formation have been discovered, including irregular spatiotemporal patterns. Here we investigate the formation of Turing patterns in simplicial complexes, where links no longer connect just pairs of nodes but can connect three or more nodes. Such higher-order interactions are emerging as a new frontier in network science, in particular describing group interaction in various sociological and biological systems, so understanding pattern formation under these conditions is of the utmost importance. We show that a canonical reaction-diffusion system defined over a simplicial complex yields Turing patterns that fundamentally differ from patterns observed in traditional networks. For example, we observe a stable distribution of Turing patterns where the fraction of nodes with reactant concentrations above the equilibrium point is exponentially related to the average degree of 2-simplexes, and we uncover parameter regions where Turing patterns will emerge only under higher-order interactions, but not under pairwise interactions.