ImPartial

ImPartial performs cell instance segmentation of noisy multiplex spatial tissue images using self-supervised multi-channel quantized imputation and complementary self-supervised objectives to reduce dense annotation requirements.

Key Features:

• Self-Supervised Multi-Channel Quantized Imputation: Augments the segmentation objective by modeling each class as a mixture of distributions through multi-channel quantized imputation, reducing reliance on perfect pixel-wise reconstruction.
• Scribbled Annotations: Leverages sparse scribbled annotations to refine and estimate pseudo-labels for segmentation training.
• Self-Supervised Denoising Parallel Task: Incorporates self-supervised denoising as a parallel task to improve segmentation performance when annotations are scarce.
• Alignment with Segmentation Goals: Optimizes a self-supervised classification objective that is explicitly aligned with instance segmentation outcomes.

Scientific Applications:

• Cancer highly-multiplexed imaging datasets: Demonstrated superior performance across various cancer datasets acquired with highly-multiplexed imaging modalities in real clinical settings.
• Spatial tissue analysis and cell instance segmentation: Applicable to noisy multiplex spatial tissue images where accurate cell instance segmentation is required for downstream analysis.

Methodology:

ImPartial uses self-supervised multi-channel quantized imputation that models classes as mixtures of distributions, utilizes scribbled annotations to estimate pseudo-labels, employs a self-supervised denoising parallel task, and optimizes a self-supervised classification objective based on the observation that perfect pixel-wise reconstruction or denoising is not necessary for accurate segmentation.