Note: Code and datasets are not publicly available due to a confidentiality agreement with Datrix S.p.A. The fluorescence microscopy dataset is proprietary clinical data from INT Milan and is not publicly available.
Internship project · Datrix S.p.A., Milan · in collaboration with Istituto Nazionale dei Tumori (INT), Milan · 2023–2024
Reza Rajaee
A reimplementation of the Noise2Void (N2V) self-supervised denoising framework (Krull et al., 2019), applied to proprietary fluorescence microscopy images provided by the Istituto Nazionale dei Tumori (INT), Milan. The implementation follows the original paper's approach, including the blind-spot masking scheme and 2D U-Net architecture, and was applied to volumetric HepG2 cell microscopy data by processing the data as 2D slices.
Krull A, Buchholz T-O, Jug F. Noise2Void — Learning Denoising from Single Noisy Images. CVPR, 2019. arXiv:1811.10980
HepG2 fluorescence microscopy (proprietary) Volumetric fluorescence microscopy images provided by the Istituto Nazionale dei Tumori (INT), Milan in Leica Image Format (.lif). Not publicly available.
Traditional denoising networks require pairs of noisy and clean images for training, which are often impossible to acquire in biomedical microscopy. N2V overcomes this by training directly on single noisy images, without any clean ground truth.
The key insight is the blind-spot network: the receptive field of each pixel excludes the pixel itself. Since noise is assumed to be pixel-wise independent, the network cannot use the center pixel to predict itself — it must instead learn to estimate the signal from surrounding context. This prevents the network from learning the identity mapping and forces it to learn meaningful denoising.
In practice, N2V implements the blind-spot using a masking scheme: during training, a randomly selected subset of pixels in each patch has its center value replaced with a randomly drawn value from the surrounding area. The loss is computed only on these masked pixels, using the original noisy values as targets.
Two statistical assumptions underlie the method:
- The signal is spatially correlated — neighbouring pixels carry information about each other
- The noise is pixel-wise conditionally independent given the signal
The microscopy data was loaded from Leica Image Format (.lif) using aicsimageio. The volumetric data was prepared for N2V training by extracting 2D patches from the image slices and generating training batches according to the N2V patch-based masking scheme, following the approach described in the original paper.
- Network: 2D U-Net with encoder-decoder structure and skip connections, following the original paper
- Training objective: MSE loss computed only on blind-spot masked pixels
- Framework: CSBDeep / Keras
- Self-supervised image denoising without clean targets
- Blind-spot networks and the pixel masking scheme
- N2V training scheme and patch-based implementation
- Fluorescence microscopy image characteristics and noise properties
- Loading and handling volumetric microscopy data in Leica Image Format
| Area | Details |
|---|---|
| Deep learning | U-Net · self-supervised training · blind-spot networks · patch-based training |
| Image denoising | Noise2Void · pixel masking · MSE loss · self-supervised learning |
| Biomedical imaging | Fluorescence microscopy · volumetric imaging · Leica Image Format (.lif) |
| Scientific Python | aicsimageio · NumPy · Keras · CSBDeep framework |
| Data preparation | Patch extraction · training batch generation from volumetric microscopy data |
- Krull A, Buchholz T-O, Jug F. Noise2Void — Learning Denoising from Single Noisy Images. CVPR, 2019.
- Lehtinen J et al. Noise2Noise: Learning image restoration without clean data. ICML, 2018. — Predecessor method
- Ronneberger O, Fischer P, Brox T. U-Net: Convolutional networks for biomedical image segmentation. MICCAI, 2015. — Base architecture
- Weigert M et al. Content-aware image restoration: Pushing the limits of fluorescence microscopy. Nature Methods, 2018. — CSBDeep framework
This repository documents the methodology and implementation approach. Code and datasets are not publicly available due to a confidentiality agreement with Datrix S.p.A. The fluorescence microscopy dataset is proprietary clinical data from INT Milan and is not publicly available.
Internship project · Datrix S.p.A., Milan · 2023–2024 · Reza Rajaee