The clinical manifestations of rectal cancer depend on the disease stage. Early-stage rectal cancer is often asymptomatic. As lesions progress, clinical symptoms emerge. Blood in the stool, the most common symptom, occurs as the stool passes by the tumor. Tumor progression causes intestinal narrowing, diarrhea, and difficult bowel movements. Chronic abdominal distension and pain can lead to malnutrition and weight loss. This may progress to emaciation and anemia. In locally advanced stages [1], tumor invasion into surrounding tissues causes corresponding symptoms. For instance, the invasion of the prostate or bladder can cause frequent urination, urgency, pain, or hematuria. The progression of colorectal cancer lesions follows this sequence: (1) normal mucosa, (2) adenomatous polyps, (3) dysplasia/high-grade neoplasia, (4) early rectal cancer (confined to the intestinal wall), (5) advanced rectal cancer (invading beyond the intestinal wall/lymphatic metastasis), and (6) distant metastasis. This subtle process gives rise to diverse lesion morphologies.

Colorectal cancer develops in the mucosa, leading to thickening of the bowel wall [2]. As the disease progresses, cancer cells invade the serosa and muscular layers. The colonoscope [3] allows direct observation of the mucosa to detect early lesions, such as ulcers, polyps, and tumors. Its disadvantage is the omission of minimal submucosal lesions or early cancers. Thus, it is insufficient for comprehensive screening or tumor staging in high-risk populations. Enhanced CT or MRI scans [4] can assess whether colon cancer invades the serosal layer. If a scan shows that the fat gap between the tumor and adjacent organs has disappeared, it suggests invasion. Due to limitations in imaging for assessing serosal invasion, some cases require the analysis of multiple scans and extensive clinical experience. For instance, a study [5] found that multi-slice 64-row, 128-layer spiral CT has a high diagnostic concordance rate (97.8%) for colon carcinoma and robust accuracy for TNM staging (T: 93.3%, N: 91.0%, M: 100%) in preoperative evaluation and identifying postoperative recurrence or metastasis. This method is a reliable and clinically valuable noninvasive diagnostic tool. However, current CT and proctoscope reports rely heavily on radiologists’ subjective expertise, which can affect diagnostic consistency. This subjectivity may also lead to inaccuracies in the final clinical diagnosis [6].

Pathological analysis is essential for cancer treatment and serves as the basis for histopathological diagnosis and classification. In modern hospitals, digital microscopy allows for high-resolution digitization of tissue sections, supporting pathological research and diagnosis. Furthermore, histopathological image analysis using AI and deep learning can identify information about colorectal lesions that is not visible to the naked eye. These techniques also facilitate the discovery of microscopic lesion information through the qualitative evaluation of colorectal morphological characteristics [7]. However, the histopathological diagnosis of colorectal cancer still faces significant challenges due to tumor heterogeneity and complexity [8].

The above factors lead to poor reproducibility in evaluating tumor pathology results. Cancer treatment requires accurate visual biomarker assessment, as the standards for tumor histopathological diagnosis are increasingly stringent [9]. Although image analysis has long existed in histopathology, its routine application is challenged by the "black box" nature of deep learning [10], image analysis methods, and slide digitization. The popularization of histopathological diagnosis is limited by the following four main research questions:(1)

Tumor heterogeneity: This reflects the significant heterogeneity between the histology and microenvironment of colorectal cancer. The distribution of colorectal tumors and cancer cells in different parts of varying patients differs, making it difficult to standardize the diagnosis, which is largely restricted by subjective factors. The EBM advocates that diagnostic and treatment decisions should integrate molecular pathology, clinical manifestations, and the latest evidence, avoiding reliance solely on traditional morphological results.

(2)

Difficulty in detecting and interpreting biomarkers: The distribution and intensity patterns of biomarkers in tumor cells or stromal tissues lack a visual interpretation of the differences between cancer cells and their microenvironments. EBM promotes XAI-based systematic discrimination tests and standards to ensure results are reasonable, comparable and practical.

(3)

Sampling error and representativeness: Inadequate sampling of colorectal pathological tissues may lead to the omission of key pathological features, affecting staging and treatment options. The EBM encourages the development of statistical models appropriate for imbalanced data and adherence to high-quality, evidence-based pathological sampling and examination processes.

(4)

The evidence underpinning model predictions: The primary challenge is distinguishing whether the model has identified causal pathways or learned spurious correlations. Without insight into its decision-making process, validating the clinical plausibility of its predictive features is nearly impossible. This opacity hinders the model's internal validation and external generalizability, as clinicians cannot confidently apply predictions without understanding the "why" behind diagnostic conclusions.

The review of computer vision studies in Evidence-based Medicine (EBM) reveals that deep learning models can accurately capture critical features that align with those identified by colorectal experts. However, the literature highlights a major limitation of current approaches: while some deep learning models achieve high predictive accuracy, this is often accomplished by relying on features outside the clinically relevant Regions of Interest (ROI). In other words, some studies show that deep learning models may focus on non-ROI areas as key features, achieving high predictive accuracy but lacking concordance with human clinical reasoning and interpretability. This reliance on non-interpretable features diminishes the clinical transparency of such models and raises concerns regarding their validity and adoption in real-world medical decision-making. To overcome these challenges, we propose the EASDnet model, which extracts diagnostically meaningful features consistent with oncologists' judgments, thereby promoting the interpretability and clinical utility of AI-driven diagnostic systems. To summarize, early screening for colorectal cancer is not about "using AI tools or citing statistical results" for diagnostic validity, but rather about using high-quality, verifiable scientific evidence to guide and evaluate the safety, effectiveness, and clinical value of AI models. Achieving deep collaboration between AI and doctors is crucial to promoting scientific, standardized, and intelligent histopathology screening for colorectal cancer.