CINA: Conditional Implicit Neural Atlas for Spatio-Temporal Representation of Fetal Brains

TL;DR

CINA is demonstrated to represent a fetal brain atlas that can be flexibly conditioned on GA and on anatomical variations like ventricular volume or degree of cortical folding, making it a suitable tool for modeling both neurotypical and pathological brains.

Abstract

We introduce a conditional implicit neural atlas (CINA) for spatio-temporal atlas generation from Magnetic Resonance Images (MRI) of the neurotypical and pathological fetal brain, that is fully independent of affine or non-rigid registration. During training, CINA learns a general representation of the fetal brain and encodes subject specific information into latent code. After training, CINA can construct a faithful atlas with tissue probability maps of the fetal brain for any gestational age (GA) and anatomical variation covered within the training domain. Thus, CINA is competent to represent both, neurotypical and pathological brains. Furthermore, a trained CINA model can be fit to brain MRI of unseen subjects via test-time optimization of the latent code. CINA can then produce probabilistic tissue maps tailored to a particular subject. We evaluate our method on a total of 198 T2 weighted MRI of normal and abnormal fetal brains from the dHCP and FeTA datasets. We demonstrate CINA's capability to represent a fetal brain atlas that can be flexibly conditioned on GA and on anatomical variations like ventricular volume or degree of cortical folding, making it a suitable tool for modeling both neurotypical and pathological brains. We quantify the fidelity of our atlas by means of tissue segmentation and age prediction and compare it to an established baseline. CINA demonstrates superior accuracy for neurotypical brains and pathological brains with ventriculomegaly. Moreover, CINA scores a mean absolute error of 0.23 weeks in fetal brain age prediction, further confirming an accurate representation of fetal brain development.

Figures (6)

• Figure 1: CINA comprises three phases. a) In the training phase, CINA learns to reconstruct 3D MRI with tissue segmentations of $N$ subjects. Each subject is assigned its unique trainable latent code $\textbf{z}$. Thus, the INR learns general domain knowledge while subject specific information is represented in $\textbf{z}$. b) CINA has learned a spatio-temporal representation of the fetal brain. To generate an atlas for time $t$, we regress a new $\textbf{z}_t$ from the trained latents $\{\textbf{z}_i\}_{i=1}^N$. Then, a single forward pass yields the atlas with tissue probability maps. c) To fit the atlas to a new subject of age $t'$, we perform test-time optimization on the MRI intensities, learning a new latent $\textbf{z}_t'$ while freezing the INR. Finally, we get the atlas and probability maps, tailored to the subject, via a forward pass while conditioning on $\textbf{z}_t'$.
• Figure 2: a) CINA and the CRL Atlas are matched with a target subject. The CRL atlas fails to adequately represent the enlarged ventricles. b+c) Interpolation on the conditioned anatomy for a fixed time point. Note, top left and right brains start to show corruptions as these extremes where never encountered during training (extrapolation).
• Figure 3: Learned representation of gestational age (GA). The first PCA component of a latent code is mapped against the corresponding subject's GA. Blue and orange represent latent codes learned during training and inference. a) A nearly perfect encoding of GA can be observed for the neurotypical population. b) For a mixed population, containing brains with ventriculomegaly, the correlation degrades. c) Explicitly conditioning (e/c) on ventricular volume improves the encoding. Note, in this setup, CINA has never seen a subject's GA.
• Figure 4: Temporal fetal brain atlas from 23 to 38 GA with corresponding probability maps of the cortical grey matter.
• Figure 5: Distribution of gestational age of the selected fetal subjects for the dHCP (left) and FeTA (right) dataset. The right histogram additionally shows the distribution (red) of the subjects by ventricular volume, normalized between 0 and 1.