Q-Net: AI QSM Diagnosis for Hereditary Hemochromatosis

• Q-Net is a two-stage deep learning framework that leverages QSM, R2*, and T1w data to differentiate HH from age-matched healthy controls, achieving up to 83.16% scan-level accuracy.
• It incorporates a CNN for slice-based feature extraction and a Bi-LSTM for sequential fusion, modeling inter-slice dependencies to improve diagnostic precision.
• The model focuses on expert-selected basal ganglia slices, demonstrating that targeted, learned representations can overcome limitations of qualitative visual assessments in iron deposition.

Searching arXiv for the specified paper and closely related work on QSM-based HH diagnosis. Q-Net is a two-stage deep learning framework for differential diagnosis of hereditary hemochromatosis (HH) versus age-matched healthy controls (HC) from brain MRI, specifically by exploiting iron-sensitive quantitative susceptibility mapping (QSM), $R2^\*$, and T1-weighted (T1w) data. It was introduced to address a clinically relevant setting in which visual inspection of iron-sensitive images is often insufficient, because both HH and normal aging can produce increased brain iron deposition, particularly in deep gray matter (DGM) nuclei. On a unique single-center cohort of 52 subjects with HH and 47 HC, Q-Net attained a scan-level accuracy of 83.16% and an image-level accuracy of 80.37%, supporting AI-driven differential diagnosis where qualitative review alone is not reliable (Zabihi et al., 2021).

1. Clinical objective and diagnostic setting

Hereditary hemochromatosis is the most common inherited disorder of systemic iron excess in people of European ancestry. Excessive systemic iron absorption can lead to deposition in multiple organs, and the study underlying Q-Net focused on evidence suggesting iron accumulation in the brain, especially in DGM nuclei. The central diagnostic problem is not merely detecting iron, but distinguishing HH-related deposition from age-related susceptibility changes in age-matched controls, a task for which QSM, $R2^\*$, and T1w images can appear deceptively similar on qualitative review (Zabihi et al., 2021).

Q-Net was designed as an AI-based differential diagnosis system rather than as a segmentation or lesion-detection model. Its stated contributions are threefold: demonstrating the feasibility of AI-driven HH-versus-HC diagnosis from multi-echo QSM/$R2^\*$/T1w data; introducing a two-stage, end-to-end model that learns discriminative features without manually segmented regions of interest; and leveraging slice-sequence context at the scan level via a bidirectional LSTM (Bi-LSTM) (Zabihi et al., 2021).

A central implication of the framework is methodological rather than purely classificatory. The model treats iron-sensitive MRI as a multimodal biomarker substrate whose informative variation is distributed across anatomy and across slices, rather than as a single-region measurement problem. This suggests a move away from manually engineered DGM ROI pipelines toward learned representations that can encode both local susceptibility patterns and scan-level context.

2. QSM, multi-echo GRE, and iron-sensitive signal modeling

Q-Net is grounded in the physics of multi-echo gradient-echo (GRE) MRI and QSM reconstruction. In the study, the complex GRE signal at echo time $t$ is modeled as

$$S(t) = S_0 \, e^{-t/T_2^*} \, e^{i\,\phi(t)}, \qquad \phi(t) = \gamma \, \Delta B \, t,$$

where $S_0$ is the initial signal, $T_2^*$ reflects dephasing from microscopic field inhomogeneities, $\gamma$ is the gyromagnetic ratio, and $\Delta B$ is the local field shift induced by susceptibility sources such as iron (Zabihi et al., 2021).

The QSM forward model relates the measured local field $b(\mathbf{r})$ to tissue susceptibility $R2^\*$0 through convolution with the dipole kernel,

$R2^\*$1

which is typically written in $R2^\*$2-space as

$R2^\*$3

with a commonly used dipole kernel

$R2^\*$4

Because $R2^\*$5 has zeros or near-zeros on the “magic angle” cone, inversion is ill-posed and requires regularization (Zabihi et al., 2021).

In the Q-Net study, susceptibility inversion used truncated $R2^\*$6-space division (TKD) with iterative refinement:

$R2^\*$7

with threshold $R2^\*$8, followed by iterative updates with iteration threshold 0.1 for four iterations (Zabihi et al., 2021).

In practical neuroimaging terms, the resulting $R2^\*$9 maps correlate with iron content, particularly in the globus pallidus, putamen, caudate, and thalamus. These DGM structures were central to HH–HC discrimination in the reported experiments. The emphasis on QSM is therefore not incidental: it is the modality in which the pathophysiological target, namely iron-sensitive susceptibility variation, is most directly encoded.

3. Cohort, acquisition, preprocessing, and input construction

The study cohort comprised 99 individuals: 52 with HH and 47 HC, recruited between February 2019 and February 2020 at a single tertiary referral center, Western University in London, Ontario, under ethics protocol 111467. Ages ranged from 20 to 85 years, with 67 males and 32 females, and HC participants were age- and sex-matched to the HH group. HH subjects were recruited without liver disease from a Gastroenterology clinic, whereas HC participants were recruited via flyers (Zabihi et al., 2021).

MRI was acquired on a 3T GE Discovery MR750 system with a 32-channel head coil using a STAGE multi-echo spoiled GRE protocol producing QSM, $R2^\*$0, T1, and spin density maps. Sequence parameters were TE1 = 5 ms, TE6 = 30 ms, $R2^\*$1TE = 5 ms, TR = 36 ms, FOV = $R2^\*$2 mm$R2^\*$3, matrix $R2^\*$4, and slice thickness 2 mm; images were interpolated to 0.43 × 0.43 mm display resolution. QSM was reconstructed for each echo and then SNR-weightedly combined across TE2–TE6 (Zabihi et al., 2021).

Preprocessing was performed in SMART v2.0. Skull stripping used BET on the second echo with threshold 0.2, erode 4, and island 2000. Phase unwrapping used 3DSRNCP. Background field removal used SHARP with threshold 0.05 and deconvolution kernel size 6. Susceptibility inversion used TKD with threshold 0.1 and iterative refinement as noted above. Because STAGE provides co-registered contrasts, no additional registration step was reported (Zabihi et al., 2021).

For learning, QSM, $R2^\*$5, and T1w were concatenated channel-wise into a 3-channel slice input analogous to RGB. Experts selected slices covering the basal ganglia, and two input configurations were evaluated: full-size images and roughly cropped basal ganglia images. No manual nuclei segmentation was performed. Training augmentation included histogram stretching, horizontal and vertical flips, random brightness/contrast, random gamma, grid distortion, shift-scale-rotate, and cropping; at test time, only histogram stretching and cropping were applied (Zabihi et al., 2021).

The design choice to rely on expert slice selection but not manual nuclei segmentation places Q-Net between fully handcrafted ROI analysis and unconstrained whole-volume modeling. A plausible implication is that the framework preserves clinically informed anatomical focus while avoiding the annotation burden of structure-level segmentation.

4. Two-stage architecture and sequential fusion

Q-Net couples image-level feature extraction to scan-level sequential modeling. Stage 1 is a slice-based CNN encoder and classifier built on ResNet18 initialized with ImageNet weights. The architecture comprises an initial $R2^\*$6 convolution, BatchNorm, ReLU, and max-pooling, followed by eight residual blocks. Each block contains two $R2^\*$7 convolutions and BatchNorm, with ReLU after the first convolution. Downsampling via $R2^\*$8 stride-2 convolution plus BatchNorm occurs before the 3rd, 5th, and 7th residual blocks, increasing channel depth from 64 to 128 to 256 to 512 while halving spatial dimensions. A global average pooling (GAP) layer reduces the final tensor to $R2^\*$9, and a fully connected layer produces image-level logits (Zabihi et al., 2021).

Stage 2 removes the Stage-1 fully connected layer and uses the frozen post-GAP feature vector from each slice as a sequential embedding. A one-layer Bi-LSTM consumes the per-slice sequence $t$0 for each scan. The LSTM cell is given by

$t$1

$t$2

$t$3

$t$4

$t$5

The bidirectional formulation processes the sequence in both forward and backward directions to capture full context (Zabihi et al., 2021).

Two fully connected heads are attached to the Bi-LSTM outputs. One head performs image-level classification of individual slices, and the other concatenates Bi-LSTM features across slices to classify the entire scan. The model therefore preserves slice-wise supervision while also learning scan-level decision rules. Binary HH-versus-HC classification uses cross-entropy loss,

$t$6

applied to the image-level and scan-level heads; no class weighting was reported (Zabihi et al., 2021).

This architecture embodies the paper’s main methodological claim: inter-slice dependencies matter. Rather than treating scan-level prediction as majority voting over independent slices, Q-Net models the basal ganglia slice stack as an ordered sequence whose contextual structure is diagnostically informative.

5. Optimization and evaluation protocol

Training used 10-fold cross-validation at the patient level, with each fold containing approximately 10 individuals. Nine folds were used for training and one for testing, with the process repeated across all folds. This patient-level partitioning is important because slice-level leakage would otherwise inflate performance in a 2D slice-based framework (Zabihi et al., 2021).

Stage 1 used transfer learning with an unfrozen ResNet18 backbone, Adam optimization, 100 epochs, batch size 64, and cosine-annealed learning rate from $t$7 to $t$8. Stochastic Weight Averaging (SWA) was applied for better generalization. Stage 2 used the frozen image embedding, Adam, 50 epochs, batch size 64, and fixed learning rate $t$9 (Zabihi et al., 2021).

Evaluation was reported over all folds using accuracy, sensitivity, specificity, F1-score, and AUC. Accuracy was defined as

$$S(t) = S_0 \, e^{-t/T_2^*} \, e^{i\,\phi(t)}, \qquad \phi(t) = \gamma \, \Delta B \, t,$$0

sensitivity as

$$S(t) = S_0 \, e^{-t/T_2^*} \, e^{i\,\phi(t)}, \qquad \phi(t) = \gamma \, \Delta B \, t,$$1

and specificity as

$$S(t) = S_0 \, e^{-t/T_2^*} \, e^{i\,\phi(t)}, \qquad \phi(t) = \gamma \, \Delta B \, t,$$2

AUC comparisons used DeLong’s test, with significance labels ranging from ns for $$S(t) = S_0 \, e^{-t/T_2^*} \, e^{i\,\phi(t)}, \qquad \phi(t) = \gamma \, \Delta B \, t,$$3 to **** for the most significant results, including $$S(t) = S_0 \, e^{-t/T_2^*} \, e^{i\,\phi(t)}, \qquad \phi(t) = \gamma \, \Delta B \, t,$$4 (Zabihi et al., 2021).

Baselines were ResNet and ResNetSWA. For scan-level comparison, these baselines used majority voting over per-slice predictions, whereas Q-Net used explicit Bi-LSTM-based scan aggregation. This establishes a direct comparison between sequential fusion and non-sequential slice aggregation.

6. Quantitative performance, ablations, and interpretability

The principal findings were that Q-Net outperformed both ResNet and ResNetSWA at both image level and scan level, and that roughly cropped basal ganglia inputs performed better than full-size images. The performance pattern supports two specific conclusions reported in the study: first, that scan-level Bi-LSTM fusion is superior to simple voting; and second, that biasing attention toward the basal ganglia improves discriminability (Zabihi et al., 2021).

Setting	ResNetSWA	Q-Net
Image-level, full-size	Acc 70.36 ± 0.20%; AUC 0.763	Acc 75.47 ± 0.80%; AUC 0.851
Image-level, cropped	Acc 71.17 ± 0.13%; AUC 0.766	Acc 80.37 ± 0.70%; AUC 0.866
Scan-level, full-size	Acc 76.89%; AUC not reported	Acc 80.00%; AUC 0.860
Scan-level, cropped	Acc 81.05%; AUC not reported	Acc 83.16%; AUC 0.875

At image level, full-size Q-Net reached accuracy 75.47 ± 0.80%, sensitivity 70.36 ± 1.77%, specificity 80.91 ± 1.54%, F1 75.43 ± 0.79%, and AUC 0.851. With cropped basal ganglia inputs, image-level Q-Net improved to accuracy 80.37 ± 0.70%, sensitivity 82.43 ± 1.76%, specificity 78.16 ± 1.14%, F1 80.35 ± 0.70%, and AUC 0.866. At scan level, full-size Q-Net achieved accuracy 80.00%, F1 79.99%, sensitivity 81.63%, specificity 78.26%, and AUC 0.860. With cropped inputs, scan-level Q-Net achieved accuracy 83.16%, F1 83.14%, sensitivity 85.71%, specificity 80.43%, and AUC 0.875 (Zabihi et al., 2021).

DeLong testing showed that Q-Net significantly outperformed both ResNet and ResNetSWA in AUC for both full-size and cropped inputs, with reported $$S(t) = S_0 \, e^{-t/T_2^*} \, e^{i\,\phi(t)}, \qquad \phi(t) = \gamma \, \Delta B \, t,$$5-values in the range $$S(t) = S_0 \, e^{-t/T_2^*} \, e^{i\,\phi(t)}, \qquad \phi(t) = \gamma \, \Delta B \, t,$$6 to $$S(t) = S_0 \, e^{-t/T_2^*} \, e^{i\,\phi(t)}, \qquad \phi(t) = \gamma \, \Delta B \, t,$$7 and significance level ****. ResNetSWA exceeded ResNet only modestly, with reported $$S(t) = S_0 \, e^{-t/T_2^*} \, e^{i\,\phi(t)}, \qquad \phi(t) = \gamma \, \Delta B \, t,$$8–0.059 (Zabihi et al., 2021).

Interpretability analysis used Class Activation Maps (CAMs), which consistently highlighted the basal ganglia in both full-size and cropped settings. This aligns with known DGM iron accumulation and supports the face validity of the learned features. Because CAM attention localized to the same anatomical regions that motivated the study’s QSM-based disease model, the interpretability results act less as a generic saliency visualization and more as a mechanistic consistency check.

7. Clinical role, limitations, reproducibility, and terminological scope

Q-Net operationalizes QSM/$$S(t) = S_0 \, e^{-t/T_2^*} \, e^{i\,\phi(t)}, \qquad \phi(t) = \gamma \, \Delta B \, t,$$9/T1w information as a decision support tool for HH-versus-HC differential diagnosis when visual inspection is insufficient. The paper suggests a role in aiding earlier identification of brain iron deposition patterns associated with HH, but it also delineates several limits on present generalizability (Zabihi et al., 2021).

The reported limitations are a modest sample size ($S_0$0), single-center acquisition, possible residual confounding from age-related iron increases despite age matching, unassessed scanner and reconstruction variability, dependence on a specific STAGE/TKD reconstruction pipeline, 2D slice modeling without explicit 3D spatial context, variability introduced by manual slice selection and approximate cropping, and the absence of class weighting or advanced domain adaptation (Zabihi et al., 2021).

Future directions explicitly proposed in the study include larger multi-center cohorts with harmonization, longitudinal HH studies, automated segmentation of DGM nuclei, examination of per-structure contributions from the globus pallidus, putamen, caudate, and thalamus, exploration of 3D CNNs and vision transformers for volumetric modeling, and domain adaptation and calibration across QSM pipelines (Zabihi et al., 2021).

Reproducibility is partial. Data are not publicly available because of ethics and consent restrictions, and code or model availability was not specified. At the same time, the paper reports acquisition parameters, the full QSM reconstruction pipeline, augmentation strategy, network architecture, training regimes, and the 10-fold cross-validation protocol in sufficient detail that the general approach is implementable on similar STAGE-derived QSM/$S_0$1/T1w data (Zabihi et al., 2021).

The name “Q-Net” is not unique in the broader machine-learning literature. In other arXiv contexts, “Q-Net” can denote a neural approximation of the action-value function $S_0$2 in reinforcement learning (Xiao et al., 2019), a query-informed few-shot medical image segmentation model (Shen et al., 2022), a quantum-enhanced tensor neural network (Konar et al., 2023), an exact-integration network for sigmoidal neural proxies (Subr, 2020), or a quantum-native sequence encoder (Day et al., 2022). In bibliographic usage, the domain qualifier is therefore essential: in this instance, Q-Net specifically denotes a QSM-based, two-stage deep neural network for differential diagnosis of brain iron deposition in hereditary hemochromatosis (Zabihi et al., 2021).