I'm a scientist, why can't I have faith?

Lan Ce was the first to start reading the manual, but the last to put it down.

"Have you figured it out yet?"

"That's about it. Let me hear what you have to say first, since you have more experience."

Lan Ce and Wei Linfeng did not argue over the decision-making position; on the contrary, Lan Ce was willing to take the initiative to assist Wei Linfeng.

Sometimes, understanding your own position can help things proceed more smoothly.

"Then let me first explain my understanding."

Wei Linfeng explained his competition strategy while referring to the competition manual.

Although there is only one computer in the answer area, and the final algorithm submission is based on this result, each participating team member can apply to the competition organizers for a development laptop.

The laptops underwent a comprehensive security check by the competition organizers. The wireless and Bluetooth modules were all physically removed, and the external interfaces could only be connected to the closed area network via network cable to eliminate the possibility of cheating.

Although the competition manual clearly states in black and white that the import of external data and devices is prohibited, the competition team's laptop can access the rack-mounted server array in the closed computer room of the competition venue, which is an offline computing power cluster, to complete the data processing.

Therefore, Wei Linfeng's thinking was very simple and clear.

First, we designed the algorithm structure and data processing strategy based on the problem.

Then he went to the answer area first to build the model architecture, since this step couldn't save time no matter what.

Lan Ce used his laptop to first complete the data processing algorithm, and then went to the answer area to access the offline dataset provided by the competition organizers for data processing, while also helping Wei Linfeng improve the architecture and conduct training.

As for Wang Dong and Su Yuran, they need to complete the interpretability design, clinical compliance verification, and ethical explanation of the algorithm based on the materials provided by the competition organizers.

Because of the clinical compliance requirements of Daxia, the competition organizers temporarily set up a small library in a conference room of the hotel.

It contains a series of materials related to the clinical requirements of Daxia, and finding what you need from it is a very demanding task.

Of course, this is also part of the competition.

After listening to Wei Linfeng's strategy, the others had no objections. As expected of a senior with rich competition experience, he came up with the best solution they could think of in a very short time, even though it was a new competition format.

The group then focused on discussing the algorithm for solving the problem.

This competition focuses on AI algorithms for accurate screening of early-stage pancreatic ductal adenocarcinoma (PDAC), with the aim of ruling out other potential risks based on patients' clinical data.

The competition organizers prepared a retrospective cohort dataset from seven top hospitals in China, Japan, and South Korea, spanning nearly ten years. The dataset contains 120,000 participants and covers all three modalities: imaging, laboratory tests, and textual medical records.

There are three review requirements:

1. Under the premise that the percentage of positive samples in the independent test set is 1.8% and the FPR (false positive rate) is less than 5%, the screening sensitivity should be greater than 90%.

2. The entire process of singleton reasoning takes less than 10 seconds on a single consumer-grade graphics card.

3. The interpretability module must comply with the latest version of the evidence-based medicine guidelines of the "Daxia Pancreatic Cancer Early Screening and Diagnosis Guidelines" and black box model output is prohibited.

Lan Ce offered his opinion on this.

"With a positivity rate of 1.8%, if a large multimodal model with hundreds of billions of parameters is used for brute-force fitting, it is very easy to collapse due to overfitting, and it is also a huge waste of time."

"I think algorithm design should be constrained by clinical priors. We should first lock in the high-dimensional feature space of pancreatic cancer as defined in the guidelines, and then talk about model training."

Wei Linfeng nodded in agreement.

"Then we can't just do simple feature splicing. We must perform cross-membrane biomarker alignment to anchor the lesion enhancement features in the images, the temporal changes in blood tests, and the high-risk factor features of the cases to the same clinical pathogenesis logic, so as to avoid noise interference between modalities."

"Interpretability, compliance, and model training are carried out simultaneously to complete the task with maximum efficiency."

The four quickly clarified their respective roles and immediately began to proceed according to plan.

Wei Linfeng went to the answer area first to start building the model framework. Instead of using the general multimodal large model like the other teams, he built a time-series fusion network customized for the early screening scenario based on his memory of previous competitions.

Lan Ce applied to the tournament organizers for a specially inspected laptop, although many modules had been removed and the storage content inside was almost negligible.

Fortunately, the laptop was pre-configured with the code environment and deep learning framework, and the competition organizers had done their part, so BlueCe could directly start building an automated processing pipeline for multimodal data based on clinical priors.

The moment Lan Ce got the laptop, the team next to him was stunned.

Holy crap, is this for real? There's such a thing?

Is it true that each team can only share one computer?

You should know that although everyone here is a coding expert, each computer only has one keyboard and can only be operated by one person at a time.

The presence of multiple computers, even just laptops, can boost the efficiency of the entire team by several levels.

Fortunately, not everyone is a fool. Soon, a team thought of the competition manual and successfully obtained their own computer.

"Holy crap, no wonder there are so many network cables for one computer. So it can even be used with a laptop?"

The already chaotic competition became even more chaotic, and only then did most teams realize the importance of carefully reading the competition manual.

But it seems it's already a bit late.

On the other side, Wang Dong and Su Yuran rushed to the small library where clinical data was stored.

"We seem to be the first one."

Su Yuran sighed as she looked at the empty little library.

Wang Dong carefully surveyed the surroundings. Besides rows of bookshelves, there was also a computer placed at the entrance of the small library.

A long network cable extends from behind the computer, its destination unknown.

"This computer should be used for information retrieval. Luckily we arrived early, otherwise, with only one electronic retrieval machine, we would have had to wait in line for a long time."

Moreover, the information seems to be unique, and the team that gets it first obviously has an advantage. For some important information, the later team can only wait for the team in front to use it up.

Without wasting any more time, the two quickly began searching for information.

What followed was a battle that was almost a contest of time and physical endurance.

As time went by, the contestants responsible for the code all began to experience physical exhaustion.

Building a model is no simple task; even the simplest algorithm can often take several hours.

It's not impossible for someone to sit in front of a screen for a whole day and night.

It was only at this point that most of the teams realized they had fallen into a trap.