Grand Challenge - AI4Life 鈣離子影像去噪競賽

在鈣離子成像中，研究人員拍到的影像幾乎一定含有大量雜訊，雜訊會讓真正的細胞活動訊號變得模糊、被誤判。鈣成像資料不是一張照片，而是空間和時間的維度。我們使用3D U-Net搭配nmODE，加強在時間維度的去噪能力。

競賽背景

• Calcium Imaging (鈣離子成像)

  • 螢光顯微鏡

  • 基因編碼鈣指示劑(Genetically Encoded Calcium Indicators, GECIs）

  • 觀察細胞中鈣濃度的變化，就能知道細胞的動作。

• AI4Life

  • 由歐盟 Horizon Europe 計畫資助的專案。專門為「光學顯微成像」與「生命科學」量身打造的「AI 工具資源庫」。

資料集

• White: Higher brightness → Neuron is active

• Black: Lower brightness → Background or inactive

• Noise

  • Photon Counting (光子計數)

  • Dark Noise / Thermal Noise

  • 去噪的目標就是讓「背景更黑（去除雜訊）」，同時讓「訊號的亮度變化（閃爍）」保持真實，不能因為去噪而讓原本該亮的點變暗。

System Overview

Temporal-Aware Solution: 3D U-Net & nmODE

• We should aware the time relation. → 3D U-Net

  • 3D → (time, height, width)

• neural memory Ordinary Differential Equation (nmODE)

  • nmODE block location → after bottleneck

  • nmODE effect → 時間連續性記憶體

ref: https://catalog.ngc.nvidia.com/orgs/nvidia/teams/dle/resources/unet_2d-industrial_tf1?version=20.06.0

評估

不同雜訊程度的去噪DEMO

sSNR, tSNR, stSNR 比較

• Signal-to-Noise Ratio (SNR)

• Spatial SNR (sSNR)

• Temporal SNR (tSNR)

• Saptio-Temporal SNR (stSNR)

• Final Score (avg stSNR)

結論

• We applied nmODE U-Net to address the temporal instability in the Calcium Imaging Denoising Contest.

• Our method nmODE U-Net is superior to CAREamics in terms of temporal stability (tSNR).

• At the highest noise level (F3), all metrics (sSNR, tSNR, and stSNR) from nmODE U-Net are better than CAREamics.

• In the future, we will submit our method to the official evaluation system to get a fair benchmark.

其他

• Challenge Site: https://ai4life-cidc25.grand-challenge.org/

• GitHub: https://github.com/jinwei-chang/grand-challenge-ai4life-cidc-2025

• Lesson report slide: https://docs.google.com/presentation/d/195icUD73Jas9jiSPeVpV9p4xOArwPccEfVC_uVx8whE/edit?usp=sharing

Reference

• [1] Yi, Z. nmODE: neural memory ordinary differential equation. Artif Intell Rev 56, 14403–14438 (2023). https://doi.org/10.1007/s10462-023-10496-2

• [2] Chen, Ricky TQ, et al. "Neural ordinary differential equations." Advances in neural information processing systems 31 (2018).