Overview and Model Performance

A comparative examination of the three architectures demonstrated a consistent performance hierarchy, with DenseNet outperforming InceptionV3 and Custom CNN. Table 1 summarises the global metrics acquired from cross-validation, including accuracy, macro-F1, macro-AUC, and their respective 95% confidence intervals.

Table 1 Summary of overall model performance (mean across folds ± 95% CI)

DenseNet outperformed InceptionV3 and the custom CNN on all macro-level metrics. In terms of accuracy, the 95% confidence interval does not overlap with the other models, showing a considerable statistical gain. The confidence intervals for macro-F1 and macro-AUC partialy overlap with InceptionV3, however DenseNet still outperforms InceptionV3 on all metrics. The custom CNN produced a moderate result but maintained stable performance, indicating the dataset's inherent complexity when trained from scratch. InceptionV3 provided an intermediate outcome, owing to its multi-scale feature extraction.

Overall, these results show that transfer learning significantly improves tauopathy progression classification, especially when adopting architectures with strong feature reuse, such as DenseNet.

Per-Class Performance

The F1 scores by class (Table 2) show a significant variation in classification difficulty between the four stages of tauopathy progression. All three structures follow a generally consistent pattern, with the 1-week class (Class 1) having the highest discriminability and the 1-day class (Class 0) the lowest.

Table 2 Mean per-class F1-scores for the three modelsEarly Stage (Class 0–1 day)

Class 0 performance was poor across all architectures (63% for CNN, 51% for InceptionV3, and 26% for DenseNet). This may be rationalised by the low AT8-positive signal observed 24 h after injury, when phosphorylated tau is sparse and morphological signs are faint (Edwards et al., 2020; Hoshino et al., 1998). The limited separability of Class 0 is biologically plausible, as early post-injury tissue closely resembles baseline neuropil, as reported in early post-injury histology in P301S/TBI models (Yoshiyama et al., 2007).

Acute Stage (Class 1–1 week)

Class 1 was the most accurately identified class, with F1 scores of 95% (DenseNet), 92% (InceptionV3), and 69% (CNN). This finding is biologically supported: one week after injury, the cortex shows substantial AT8-positive perisomatic aggregates and numerous neuropil threads, resulting in a robust and spatially coherent histopathological signature (Edwards et al., 2020; Yoshiyama et al., 2007). Consistent with this, several TBI models show a significant rise in tau abnormalities throughout this interval (Tran et al., 2011). As a result, DenseNet obtained a very good separability for this class.

Intermediate Stage (Class 2–1 month)

Class 2 performed averagely (55% for CNN, 54% for InceptionV3, and 63% for DenseNet). Misclassifications were most common with Class 3, which is biologically compatible with the gradual, rather than sudden, increase of tau pathology one to three months after injury (Clavaguera et al., 2009; Edwards et al., 2020; Yoshiyama et al., 2007). AT8 positive tau is certainly present in this intermediate stage, but its distribution is heterogeneous and spatially variable, resulting in less stable morphological cues and contributing to inconsistent predictions (Arendt et al., 2016; Edwards et al., 2020; Tran et al., 2011).

Late Stage (Class 3–3 months)

F1-scores for Class 3 varied from 81% (CNN) to 74% (InceptionV3) and 71% (DenseNet), indicating moderate discriminability. This behaviour reflects the underlying biology: by three months, tauopathy is widespread but still morphologically overlapping with the disease reported at one month, indicating a mostly quantitative rather than qualitative development (Clavaguera et al., 2009; Edwards et al., 2020). Unlike the extremely identifiable AT8-positive patterns found after one-week, later stages lack unambiguous class-specific structural motifs, and their heterogeneous distribution lowers the confidence of CNN-based predictions (Arendt et al., 2016; Yoshiyama et al., 2007).

When examined together, the per-class performance reveals a consistent biological trend. The highest separability was seen at one week, which corresponds to the time when AT8-positive tau pathology becomes most prominent and morphologically distinct (Edwards et al., 2020; Tran et al., 2011; Yoshiyama et al., 2007). In contrast, the lowest separability was seen after one day, which is consistent with the near absence of phosphorylated tau following injury (Edwards et al., 2020; Hoshino et al., 1998; Tran et al., 2011). The intermediate difficulty in separating 1-month and 3-month samples reflects the gradual and mainly quantitative evolution of tauopathy in later stages, when At8-positive inclusions increase in burden while sharing structural features (Arendt et al., 2016; Clavaguera et al., 2009; Edwards et al., 2020). These findings suggest that the models accurately captured important features of the temporal dynamics of tau accumulation over post-injury intervals.

Confusion Matrixes and Class Separability

The confusion matrices (Fig. 3 (a), (b) and (c)) for the three architectures (CNN, InceptionV3, and DenseNet) show consistent misclassification patterns which follow the biological characteristics of tau pathology across time, particularly for transfer-learning models. Although the total number of correct predictions varies amongst models, the underlying error pattern is similar, with CNN exhibiting more architecture-driven confusion patterns, likely due to its limited representational capacity.

Fig. 3

Comparative performance of: (a–c) Confusion matrices for the CNN, InceptionV3, and DenseNet models, respectively; and (d–f) Corresponding class-wise ROC curves (with AUC values) for the CNN, InceptionV3, and DenseNet models

In both transfer models (InceptionV3 and DenseNet), Class 1 (1 week) demonstrated the highest separability, with strongly concentrated diagonal entries and minimal cross-class confusion. This aligns with the marked increase in AT8-positive perisomatic aggregates and neuropil threads typically observed at this temporal stage (Edwards et al., 2020). In contrast, the baseline CNN did not capture these discriminative features as efficiently, demonstrating higher confusion among all classes, including 1 week.

In contrast, Class 0 (1 day) is inconsistently recognised. The CNN model has the highest true-positive count (554), whereas InceptionV3 and DenseNet have a higher distribution of Class 0 data into Classes 2 and 3. This reflects the naturally weak and diffuse AT8 signal at 24 h, which lacks the structured morphology that deeper models expect. As a result, the DenseNet and InceptionV3 models, which depend more heavily on prominent and multi-scale patterns, frequently misread weak early stage staining as more advanced stages (most often Class 2).

Confusion between Classes 2 (1 month) and 3 (3 months) exists in both transfer-learning models, although in an asymmetrical manner: DenseNet misclassifies some 1-month images as 3-month images, whereas InceptionV3 exhibits the opposite trend. In the baseline CNN model, the confusion is minimal. This behaviour is consistent with the biological similarity between the two stages, since both show extensive AT8-positive neuritic degeneration, with the differences being mostly quantitative rather than qualitative (Edwards et al., 2020; Yoshiyama et al., 2007).

DenseNet displayed strong diagonal performance, especially in Class 3 (3 months), showing strong separability for late-stage pathology. InceptionV3 performs comparably, with the maximum correct classification count for Class 1 (1 week), but with slightly higher dispersion in Classes 2 and 3. In contrast, the custom CNN shows the most diverse confusion patterns, particularly between the early (0–1) and advanced (2–3) stages.

Overall, the confusion matrices show that much of the difficulty in classification is due to the underlying biological progression of tauopathy: a weak or non-existent AT8 signal after 1 day (Edwards et al., 2020; Hoshino et al., 1998; Tran et al., 2011), a clear morphological peak after 1 week driven by dense perisomatic and neuropil AT8-positive pathology (Edwards et al., 2020; Yoshiyama et al., 2007), and a more gradual quantitative difference between 1 and 3 months (Clavaguera et al., 2009; Edwards et al., 2020). This trend is more visible in transfer learning models (InceptionV3 and DenseNet), where misclassifications correlate with the predicted temporal structure of tau aggregation. However, the basic CNN model's insufficient ability to extract subtle mid- to late-stage features cause certain non-biologically driven errors.

ROC Curve Analysis

The ROC curves (Fig. 3 (d), (e) and (f)) provide additional insight into class separability across the three architectures. Consistent with previous analyses, all models show that Class 1 (1 week) and Class 3 (3 months) have higher AUC values due to their prominent AT8-positive morphological fingerprints (Edwards et al., 2020; Tran et al., 2011; Yoshiyama et al., 2007). DenseNet and InceptionV3 have comparable ROC curve profiles, with DenseNet achieving somewhat better AUCs at the later stages and InceptionV3 at the most pronounced stage (1 week).

Class 0 (1 day) has the lowest AUC value among transfer learning models (InceptionV3: 0.78, DenseNet: 0.79), whereas the CNN model has the lowest value for Class 2 (0.80 vs 0.84 for Class 0). Nonetheless, this is consistent with the low AT8 immunoreactivity observed after 24 h of injury, where tau phosphorylation is weak or absent (Edwards et al., 2020; Hoshino et al., 1998; Tran et al., 2011).

In intermediate-stage tauopathy (Class 2—1 month), all architectures show moderate AUC values (0.79–0.82), indicating progressive but partially overlapping tau deposition (Clavaguera et al., 2009; Edwards et al., 2020).

Overall, ROC patterns show that separability increases when tau pathology is either strongly expressed (1 week, 3 months) or has a distinct structural organisation. These patterns are qualitatively consistent with the reported temporal dynamics of AT8-positive tau accumulation following TBI and in P301S tauopathy models.

Comparative Statistical Analysis

To compare the three architectures, accuracy and macro-F1 values were analysed across the cross-validation folds for each model. DenseNet consistently achieved higher fold-level values compared to InceptionV3 and custom CNN, which is consistent with the global metrics reported in Sects. 3.1 and 3.2. The 95% confidence intervals calculated from the fold distributions indicate minimal overlap between DenseNet and the other models, suggesting a meaningful performance advantage with the internal validation setting.

Although no formal statistical test was conducted, the fold-wise distributions show a clear separation between DenseNet and the other architectures, whereas CNN and InceptionV3 have more similar variability. These descriptive trends corroborate DenseNet's superior stability and discriminative capabilities.

Importantly, these comparisons represent relative internal performance rather than conclusive evidence or overall superiority. Since neither animal-level IDs or external validation were provided, fold-level estimates could capture biological or acquisition-related variability. Extensive external testing will be required to establish whether the reported performance hierarchy holds true across diverse staining intensity, imaging settings, and biological samples.

Neurobiological Contextualization of Model BehaviourInterpretation of Classification Behaviour Across Temporal Stages

The per-class performance patterns reported across the three architectures substantially are consistent with documented biological aspects of tauopathy development after traumatic brain injury. The 1-week class (Class 1) has high discriminability, which corresponds to the typical increase of AT8-positive immunoreactivity seen at this stage. Around one week after injury, hyperphosphorylated tau accumulates in perisomatic aggregates, neuropil threads, and discrete laminar patterns, particularly in cortical layers II/III (Arendt et al., 2016; Edwards et al., 2020; Yoshiyama et al., 2007). These well-defined and spatially consistent morphological signals provide a strong visual signature, allowing deep learning models, particularly DenseNet and InceptionV3, to attain F1 scores of around or above 0.90 in this class.

In contrast, Class 0 (1 day) was consistently challenging across models. At 24 h post-injury, AT8-positive tau is usually sparse or absent, with only subtle background staining and little to no organised deposits or neuritic pathology (Edwards et al., 2020; Hoshino et al., 1998; Tran et al., 2011). As a result, numerous images resemble non-pathological cortex. This weak signal explains why all architectures struggle to differentiate this class, and why deeper, more regularised DenseNet has particularly low sensitivity: deeper architectures prioritise more prominent and higher-order patterns, potentially underfitting extremely small morphological cues.

The intermediate performance of Class 2 (1 month) and the confusion shared with Class 3 (3 months) demonstrate the gradual and continuous nature of tau accumulation in later stages. Between one and three months after injury, AT8-positive pathology becomes more pervasive, although the changes are primarily quantitative rather than qualitative: both time periods show severe neuritic degeneration and diffuse cortical tau deposition (Clavaguera et al., 2009; Edwards et al., 2020; Jucker & Walker, 2011). This results in overlapping morphological profiles, and all models achieve mid-range F1 scores for these stages, with frequent misclassifications between them, especially in transfer-learning models.

When combined, the categorisation behaviour of the three models resembles important neuropathological trajectories of tauopathy rather than arbitrary error patterns. Rather than spurious artefacts, the models appear to rely on meaningful biological features, such as strong, focal AT8 immunoreactivity at one week, weak or absent early signal at one day, and progressive yet overlapping pathology at later stages. This adds to the biological plausibility of the learnt representations and shows that many model errors occur at periods where the underlying pathology is itself difficult to distinguish, even for experienced observers.

Mechanistic Links between Traumatic Brain Injury, Microbleeds, Iron Toxicity and Tau Aggregation

Traumatic brain injury initiates a series of secondary pathological processes that extend well beyond the original trauma mechanism. Among the most significant are microvascular disruption, microbleeds, and iron deposition, which create a biochemical environment that promotes tau phosphorylation, aggregation, and neurodegeneration over time. Blood–brain barrier disruption and microhaemorrhages release haemoglobin, heme, and iron into the parenchyma, causing oxidative stress, mitochondrial dysfunction, and chronic microglial activation, according to experimental and clinical research (Christodulou et al., 2025a). Excess iron catalyses free radical production and increases neuronal vulnerability, which exacerbates the neurotoxic effects of phosphorylated tau, accelerating synaptic dysfunction and cortical atrophy.

These haemorrhage-associated processes develop slowly, creating a progressive inflammatory and iron-rich environment that lasts for weeks to months after TBI. Chronic iron-driven inflammation is becoming recognised as the mechanistic link between TBI and long-term tauopathy, with microbleeds and hemosiderin deposits acting as focal initiators of persistent neurogenerative signalling (Christodulou et al., 2025b). This framework is consistent with observations in repetitive or blast related TBI, in which iron buildup and microvascular pathology worsen tau seeding, aggregation, and cortical degeneration (Christodulou et al., 2025c).

Within this mechanistic context, our models' classification behaviour can be biologically interpretable. The low separability at 1 day is most likely due to the fact that microbleed-related iron toxicity and inflammatory cascades have only recently begun, with no detectable AT8-positive aggregates. The strong discriminability at 1 week corresponds to the onset of significant tau phosphorylation, which is consistent with the early effects of iron-mediated oxidative stress. Finally, the partial overlap between the 1- and 3-month stages shows the chronic, gradual nature of iron-associated neurodegeneration, in which tau accumulation increases but remains morphologically continuous rather than sharply distinguished.

When considered together, microbleeds, iron deposition, and inflammation could provide a mechanistic explanation for both the biological trajectory of tauopathy after TBI and the class-specific patterns observed in our deep learning models.

Error Analysis

A comparison of misclassification patterns across all models reveals that the errors were not random, but rather followed biologically relevant tendencies. The most common source of errors was confusion between one month (Class 2) and three months (Class 3), which is consistent with the progressive and continuous nature of tau aggregation in P301S tauopathy models. Both stages show widespread AT8-positive neuritic degeneration, with more quantitative than qualitative differences (Clavaguera et al., 2009; Edwards et al., 2020; Jucker & Walker, 2011). As a result, even the most regularised architectures, such as DenseNet, exhibited residual ambiguity between these two classes.

A second recurring pattern was that transfer-learning models, particularly DenseNet, misclassified early-stage images (Class 0—1 day) as later stages, most commonly Class 2 or Class 3 in the transfer-learning models. This behaviour is biologically plausible: 24 h after damage, AT8 immunoreactivity is weak or non-existent, resulting in images that closely resemble the baseline cortex (Edwards et al., 2020; Hoshino et al., 1998; Tran et al., 2011). Deep architectures that prioritise strong and coherent morphological characteristics may thus underfit the highly delicate cues available at this time period, resulting in a systematic overestimation of disease progression.

In contrast, errors involving Class 1 (1 week) were uncommon, especially in transfer learning models. This class has both strong F1-scores (except for the CNN model) and steep ROC curves, with misclassification occurring primarily in borderline cases with unusually low AT8 signal. This finding confirms that model performance is directly related to the intrinsic morphological expressiveness of each pathological stage.

Overall, the observed error patterns suggest that model performance is influenced more by the biological structure of the dataset than by architectural differences. Misclassifications occurred most frequently when the underlying pathology was either (i) too subtle to reliably distinguish from baseline tissue (Class 0) or (ii) too continuous to be accurately separated (Classes 2 and 3). These findings indicate that, rather than purely architectural improvements, additional contextual information, such as patch aggregation, slide-level metadata, or multimodal inputs, will be required to achieve further performance gains.

Comparison with Related Deep Learning Approaches in Tau Pathology

Although no previous research has focused on multi-stage temporal classification of TBI-induced tauopathy using AT8-stained histology, our findings can be contextualised within relevant deep-learning approaches in tau pathology.

Signaevsky et al. (Signaevsky et al., 2019) and Tang et al. (Tang et al., 2019) demonstrated that CNNs may quantify tau burden and accurately distinguish Alzheimer's disease pathological subtypes, however these models relied on static snapshots rather than temporally structured progression. Similarly, Wang et al. (Wang et al., 2024) demonstrated that CNNs are effective at capturing fine cortical microarchitecture, while Zou et al. (Zou et al., 2021) found that tau-PET diagnostic value increased with deep learning.

Compared to existing approaches, our framework addresses more subtle morphological differences and a more difficult four-stage classification problem. The present study's performance hierarchy (DenseNet > InceptionV3 > Custom CNN) is consistent with previous findings that connected or multiscale networks outperform in small-sample histopathology tasks (Hussain et al., 2019; Raghu et al., 2019; Shin et al., 2016). Thus, while conceptually similar to existing tau-focused models, our study goes beyond them by demonstrating that deep learning can track the temporal trajectory of tau accumulation after TBI, a task not